Problem Statements Analysis

From what to build to how to build it — our platform helps you understand, plan, and solve SIH 2026 problem statements with complete insights into requirements, solution approaches, technology stacks, architecture, AI/ML, and implementation roadmaps.

Scroll down
01 / SIH 2026

PS: SIH26187

AI-Based Intelligent Video Analytics Platform for Border Surveillance using existing CCTV Infrastructure.

Description: • Background Border security forces deploy CCTV cameras at Border Out Posts(BOPs), check posts, border roads, and other strategic locations for surveillance and monitoring. However, conventional CCTV systems primarily provide video recording and live monitoring capabilities,requiring continuous human observation. Advanced surveillance functionalities such as Facial Recognition Systems (FRS), Automatic Number Plate Recognition (ANPR), intrusion detection, and object tracking often require specialized hardware and proprietary solutions,making large-scale deployment costly and difficult, particularly in remote border areas. • Description The proposed solution aims to develop an AI-driven software platform capable of transforming existing CCTV infrastructure into an intelligent surveillance network without requiring dedicated FRS, ANPR, or smart-camera hardware. The platform shall ingest live video streams from standard IP-based CCTV cameras and perform real-time video analytics using Artificial Intelligence and Computer Vision techniques. The solution should provide capabilities such as: • Human detection and tracking • Vehicle detection and classification • Face detection • Automatic Number Plate Recognition (ANPR) • Virtual fence intrusion detection • Suspicious activity detection • Night-time movement detection • Real-time alert generation and event logging • Expected Solution The proposed system should leverage Artificial Intelligence, Machine Learning, Computer Vision, and Video Analytics to create a software-defined surveillance platform capable of extracting actionable intelligence from existing CCTV infrastructure. The solution should: • Eliminate dependence on expensive dedicated surveillance hardware. • Enable intelligent monitoring through AI-powered video analytics. • Provide real-time alerts for security incidents and border intrusions. • Support facial recognition, vehicle identification, and behavioral analytics through software. • Improve situational awareness and response time for border security forces. • Support integration with existing command and control systems. • The final solution should be cost-effective, scalable, and suitable for deployment across remote border locations and strategic installations. • Possible Project Name IBVAP – Intelligent Border Video Analytics Platform
Organization: Ministry of Home Affairs
Department: Sashastra Seema Bal (SSB), Police II Division
Category: Software
Theme: Blockchain & Cybersecuritytd>
Youtube Link: Not provided
Dataset Link: Not provided
Contact info: Not provided

1. Executive Snapshot

The proposed project, IBVAP (Intelligent Border Video Analytics Platform), addresses a critical operational gap facing the Sashastra Seema Bal (SSB) along India's open borders with Nepal and Bhutan [PS, RESEARCHED]. Border security forces currently rely on legacy IP CCTV networks deployed across Border Out Posts (BOPs), check posts, and border roads [PS]. However, these conventional systems require continuous human monitoring and lack automated threat intelligence [PS]. Upgrading to hardware-dependent AI cameras or specialized proprietary video processing hardware across thousands of remote locations is economically prohibitive and operationally complex [PS, INFERENCE]. IBVAP delivers a software-defined, edge-hybrid video analytics layer that ingests standard RTSP streams from existing CCTV cameras, performing multi-stream real-time AI analytics (human tracking, vehicle classification, face recognition, ANPR, virtual fence intrusion, suspicious behavior, and night-time motion detection) directly on localized edge nodes with full store-and-forward offline resilience [PS, ASSUMPTION].

Metadata Summary

Parameter Value Evidence
Problem Statement ID SIH26187 [PS]
Title AI-Based Intelligent Video Analytics Platform for Border Surveillance using existing CCTV Infrastructure [PS]
Organization Ministry of Home Affairs (MHA) [PS]
Department Sashastra Seema Bal (SSB), Police II Division [PS]
Category Software [PS]
Theme Blockchain & Cybersecurity [PS]
Deadline 30 September 2026 [PS]
Ideas Counter 1/500 [PS]
Dataset Link None Provided [PS]
Contact / Media None Provided [PS]

Fundamental System Objective: Transform legacy IP CCTV infrastructure at remote border outposts into an automated, edge-intelligent surveillance network using software-defined Computer Vision pipelines that generate real-time alerts without requiring hardware replacement [PS, INFERENCE].

2. Problem Deconstruction

The problem statement has been broken down into atomic, traceable explicit requirements alongside derived operational implications:

  1. Software-Defined Architecture on Legacy Infrastructure

    • Source Text: "transforming existing CCTV infrastructure into an intelligent surveillance network without requiring dedicated FRS, ANPR, or smart-camera hardware." [PS]
    • Interpretation: The platform must interface directly with standard RTSP/ONVIF video streams provided by non-AI IP cameras [INFERENCE].
    • Implementation Implication: Processing must take place on localized Edge Gateways or regional C2 servers without modifying camera firmware or adding physical camera modules [INFERENCE].
    • Confidence: [PS]

  2. RTSP Live Video Stream Ingestion

    • Source Text: "ingest live video streams from standard IP-based CCTV cameras and perform real-time video analytics..." [PS]
    • Interpretation: High-throughput streaming protocol support with automatic reconnect, frame-buffering, and multi-resolution handling [INFERENCE].
    • Implementation Implication: Multi-threaded RTSP ingest engine (e.g., MediaMTX/GStreamer) capable of decoding H.264/H.265 video streams [ASSUMPTION].
    • Confidence: [PS]

  3. Human Detection and Multi-Object Tracking

    • Source Text: "Human detection and tracking" [PS]
    • Interpretation: Continuous tracking of individual persons across video frames to analyze trajectory and dwell time [INFERENCE].
    • Implementation Implication: Object detection model (YOLOv8/v10/v11) coupled with a multi-object tracking algorithm (ByteTrack / BoT-SORT) [ASSUMPTION].
    • Confidence: [PS]

  4. Vehicle Detection and Classification

    • Source Text: "Vehicle detection and classification" [PS]
    • Interpretation: Categorize moving objects into classes (e.g., Sedan, SUV, Truck, Bus, Two-Wheeler, Auto-rickshaw) [INFERENCE].
    • Implementation Implication: Multi-class bounding box classification model optimized for Indian vehicle profiles [ASSUMPTION].
    • Confidence: [PS]

  5. Face Detection and Recognition (FRS)

    • Source Text: "Face detection... Support facial recognition... through software." [PS]
    • Interpretation: Extract human facial regions from video streams and match them against a local or central watchlist of interest [INFERENCE].
    • Implementation Implication: Two-stage biometric pipeline (RetinaFace/SCRFD for detection + ArcFace embeddings matched against Milvus/FAISS vector index) [ASSUMPTION].
    • Confidence: [PS]

  6. Automatic Number Plate Recognition (ANPR)

    • Source Text: "Automatic Number Plate Recognition (ANPR)... vehicle identification" [PS]
    • Interpretation: Detect registration plates and perform Optical Character Recognition (OCR) under varied lighting and angles [INFERENCE].
    • Implementation Implication: Plate localization model paired with an OCR reader fine-tuned for Indian High Security Registration Plates (HSRP) and state formats [RESEARCHED, ASSUMPTION].
    • Confidence: [PS]

  7. Virtual Fence Intrusion Detection

    • Source Text: "Virtual fence intrusion detection" [PS]
    • Interpretation: Operator-defined tripwires and polygon geofences that trigger instantaneous alerts when crossed by humans or vehicles [INFERENCE].
    • Implementation Implication: Polygon-bounding spatial intersection check using vector math on bounding box base coordinates [ASSUMPTION].
    • Confidence: [PS]

  8. Suspicious Activity and Behavioral Analytics

    • Source Text: "Suspicious activity detection... behavioral analytics through software." [PS]
    • Interpretation: Detect anomalous behaviors such as loitering, abrupt running, abandoned baggage, or unauthorized group formation near perimeter lines [INFERENCE].
    • Implementation Implication: Dwell-time calculation, velocity vectors, and temporal action recognition algorithms operating on tracked bounding box trajectories [INFERENCE].
    • Confidence: [PS]

  9. Night-time Movement Detection

    • Source Text: "Night-time movement detection" [PS]
    • Interpretation: Detect motion and objects in extreme low-light or Infrared (IR) night-vision modes standard on legacy cameras [INFERENCE].
    • Implementation Implication: Pre-processing enhancement pipeline (CLAHE / Gamma correction / Zero-DCE) fed into low-light fine-tuned vision models [ASSUMPTION].
    • Confidence: [PS]

  10. Real-time Alerting, Logging, and Command & Control Integration

    • Source Text: "Real-time alert generation and event logging... Support integration with existing command and control systems." [PS]
    • Interpretation: Sub-second alert generation with video snapshot/clip attachment, structured audit logging, and open API push to existing MHA C2 platforms [INFERENCE].
    • Implementation Implication: MQTT/WebSocket messaging broker and REST/CoT (Cursor on Target) integration endpoints [ASSUMPTION].
    • Confidence: [PS]

  11. Remote Border Deployment & Cost Effectiveness

    • Source Text: "cost-effective, scalable, and suitable for deployment across remote border locations and strategic installations." [PS]
    • Interpretation: Operational capability under constrained network bandwidth, edge compute constraints, and harsh physical environments [INFERENCE].
    • Implementation Implication: Local edge processing architecture minimizing WAN bandwidth usage (pushed alerts only, not full streams) with offline store-and-forward capability [INFERENCE].
    • Confidence: [PS]

3. Stakeholders & User Needs

Stakeholder Goal Current Pain System Need
BOP Duty Officer / Patrol Commander Monitor 10–30 live camera feeds simultaneously and respond rapidly to border intrusions [KNOWN]. Screen fatigue; critical events missed during manual visual checks; lack of actionable real-time alerts [PS]. Automated desktop/mobile alert triggers with snapshot, exact camera position, intruder classification, and audio-visual alarms [INFERENCE].
Quick Response Team (QRT) / Field Patrol Intercept unauthorized border crossers or suspect vehicles before they clear the border zone [KNOWN]. Delayed notification; lack of physical visual description or vehicle license plate details prior to interception [INFERENCE]. Mobile companion app / push alerts with exact GPS coordinates of intrusion, target photo, and vehicle plate number [ASSUMPTION].
Sector HQ Command Center Analyst Maintain macro situational awareness across 50+ BOPs and cross-reference criminal watchlists [RESEARCHED]. Isolated video feeds; inability to search historical video by vehicle plate or face match across multiple outposts [INFERENCE]. Centralized GIS dashboard showing multi-BOP threat maps, unified search across vehicle/face databases, and high-level analytical dashboards [INFERENCE].
MHA / SSB Systems Administrator Maintain system security, manage software deployments, and integrate with legacy government databases [KNOWN]. High bandwidth costs if streaming raw video to cloud; vendor lock-in with proprietary camera brands [PS, INFERENCE]. Hardware-agnostic software platform, low-bandwidth edge architecture, localized Docker deployments, and open API/ONVIF standards compliance [ASSUMPTION].

4. Requirements Traceability

Feature / Architecture Component Explicit Requirement Justification Derived Operational Need Evidence Tag
RTSP / ONVIF Multi-Stream Ingestion Engine "ingest live video streams from standard IP-based CCTV cameras" Support mixed vendor IP cameras (Hikvision, Dahua, CP Plus, Axis) without modifying physical setup. [PS]
YOLOv11 Person & Vehicle Detector "Human detection and tracking... Vehicle detection and classification" High-accuracy detection of pedestrians, cars, trucks, bikes under varied weather and illumination. [PS]
ByteTrack Object Tracking Module "Human detection and tracking" Maintain unique IDs across frame occlusions for movement trajectory estimation. [PS]
RetinaFace + ArcFace Vector Search Engine "Face detection... Support facial recognition" Extract facial features and match against a localized database of POIs (Persons of Interest). [PS]
PaddleOCR / Custom HSRP OCR Pipeline "Automatic Number Plate Recognition (ANPR)" Parse standard Indian High Security Registration Plates (HSRP) and non-standard vehicle plates. [PS, RESEARCHED]
Interactive Polygon Virtual Fencing UI "Virtual fence intrusion detection" Allow duty officers to draw dynamic restricted boundaries on live camera feeds. [PS]
Temporal Loitering & Velocity Anomaly Module "Suspicious activity detection... behavioral analytics" Trigger alerts when individuals dwell in restricted zones beyond threshold limits or move at unusual speed. [PS]
Low-Light / IR Contrast Enhancement Pipeline "Night-time movement detection" Process black-and-white IR feed frames from standard cameras during night hours without hardware upgrades. [PS]
MQTT Light-Weight Alert Push & Event Storage "Real-time alert generation and event logging... remote border locations" Transmit critical alerts over low-bandwidth (2G/3G/4G/VSAT) links while caching events locally during outages. [PS, INFERENCE]
REST / Webhook / CoT Integration Gateway "Support integration with existing command and control systems" Expose standardized event outputs for external MHA/SSB C2 platforms. [PS]

5. Constraints & Environment Realities

  1. Geographic & Physical Environment: SSB guards the 1,751 km Indo-Nepal and 699 km Indo-Bhutan borders, spanning dense Terai forests, riverine plains, and high-altitude Himalayan terrain [RESEARCHED]. BOPs are situated in harsh physical conditions with extreme temperature ranges, dust, heavy monsoon rain, and winter fog [KNOWN].
  2. Connectivity Limitations: Remote BOPs frequently rely on low-bandwidth satellite (VSAT), cellular 2G/3G/4G with high latency, or unreliable line-of-sight wireless links [KNOWN]. Continuous streaming of raw 1080p video (requiring 2–4 Mbps per camera) to a centralized cloud is technically infeasible [INFERENCE].
  3. Power Grid Instability: Remote border outposts experience frequent power outages, relying on diesel generators and solar battery backups [KNOWN]. Edge compute hardware must be power-efficient (e.g., low-TDP edge servers or industrial NVIDIA Jetson Orin modules operating under 60W) [ASSUMPTION].
  4. Existing Hardware Heterogeneity: Legacy CCTV infrastructure comprises various brands (CP Plus, Dahua, Hikvision, Bosch) with differing resolutions (720p, 1080p, 4K), frame rates (10–30 FPS), codecs (H.264, H.265), and infrared illuminator ranges [KNOWN].
  5. Human Factors & Operator Literacy: Field personnel at BOPs may have limited technical IT expertise [KNOWN]. The user interface must be clutter-free, highly visual, localized, and require minimal manual setup [ASSUMPTION].

6. Research Questions

  1. How can low-bitrate RTSP streams from older IP cameras be processed for ANPR without causing severe frame drop or high processing latency?
  2. What open-source model optimization techniques (TensorRT quantization, FP16/INT8) yield sub-50ms inference time on edge hardware for 8–16 simultaneous camera streams?
  3. What is the legal and regulatory framework governing facial recognition biometrics (DPDP Act 2023) when deployed by Central Armed Police Forces (CAPF) along open international borders? [RESEARCHED]
  4. How do high-humidity monsoon conditions and fog affect infrared (IR) night-vision analytics, and can digital pre-processing filters compensate for poor visual clarity?
  5. What protocol (e.g., MQTT, NATS, WebSockets, Cursor on Target) is most resilient for synchronizing edge alert events with Sector HQ across lossy satellite links?

7. Data Requirements & Sources

Data Type Purpose Ideal Real-World Source Hackathon Substitute Access Method Availability Privacy Sensitivity Confidence
Border Surveillance RTSP Feeds Training/testing human, vehicle, and night detection in border contexts. Live SSB CCTV stream feeds at BOPs [PS]. Public surveillance datasets (e.g., VIRAT, UCF-QA, PETS2009, custom recorded RTSP video loops). Local file storage / RTSP simulation server (MediaMTX) [ASSUMPTION]. Openly Available (Public Benchmarks) Public Domain / Non-sensitive [RESEARCHED]
Indian Vehicle & License Plate Images Training ANPR and vehicle classification models. Ministry of Road Transport & Highways (MoRTH) / VAHAN database [RESEARCHED]. Public Indian License Plate datasets (e.g., Kaggle Indian License Plates, Roboflow HSRP datasets). Direct download / local disk [ASSUMPTION]. Openly Available Medium (License plate numbers) [RESEARCHED]
Watchlist Facial Biometrics Evaluating Face Detection & Recognition against target lists. MHA Criminal / POI Watchlist Database [INFERENCE]. Synthetic face generation (StyleGAN) + open face datasets (LFW, CelebA, Labeled Faces in the Wild). Local database indexing into FAISS vector store [ASSUMPTION]. Synthetically Generated High (Biometric data requires encryption & access logging) [ASSUMPTION]
Geospatial & Vector Fence Data Storing camera positions, field of view (FOV), and geofence polygons. SSB GIS Mapping / Survey of India maps [RESEARCHED]. OpenStreetMap (OSM) / GeoJSON mock coordinates for border outposts. Local PostGIS / GeoJSON files [ASSUMPTION]. Openly Available Low [KNOWN]

Data Categorization Breakdown

  • A. Explicitly Mentioned: Live video streams from standard IP CCTV cameras [PS].
  • B. Required by Inference: Vector watchlists for facial recognition, registered vehicle blacklists for ANPR, spatial GeoJSON polygons for virtual fencing, historical incident logs [INFERENCE].
  • C. Data That Must Be Collected: Real-world test video clips capturing low-light night movement, blurry license plates, and multi-person occlusions [INFERENCE].
  • D. Data Synthetically Generated: Mock criminal face embeddings, simulated vehicle license plates, synthetic alert event logs for stress testing [ASSUMPTION].
  • E. Data That Still Needs Validation: Real-world performance of low-cost CCTV IR sensors at distance >50 meters under dense fog conditions [NEEDS VALIDATION].

8. Existing Solutions & Gap Analysis

Existing System / Solution Class Description Strengths Weaknesses Relevance Gap Remaining in Current Market Source Confidence
Commercial Enterprise VMS (e.g., Milestone XProtect, Genetec Omnicast) Proprietary video management software with add-on AI plugins. Highly stable, broad camera protocol support, enterprise UI. Extremely expensive recurring licensing fees; heavy reliance on high-end server hardware; closed ecosystem [INFERENCE]. High Requires proprietary licensing per channel and high bandwidth connection to central servers; poorly adapted to low-bandwidth edge deployment [INFERENCE]. Industry Benchmark Docs [KNOWN]
Smart AI-Enabled Cameras (e.g., Hikvision AcuSense, Dahua WizSense) Cameras with embedded edge AI hardware chips. Low latency, no separate edge server needed at small scale. Requires full replacement of existing camera infrastructure; limited to fixed vendor models; restricted algorithm updates [PS]. High Fails problem requirement to utilize existing legacy CCTV infrastructure without replacing hardware [PS]. Vendor Technical Datasheets [KNOWN]
Open-Source Edge Surveillance (e.g., Frigate NVR) Lightweight local NVR using Coral TPU / OpenVINO for object detection. Free, open-source, low resource utilization, privacy-focused. Lacks native HSRP ANPR, high-capacity facial recognition pipelines, multi-BOP C2 aggregation, and defense-grade RBAC/audit trails [INFERENCE]. Medium Designed primarily for smart home/smb security; lacks border-specific behavioral analytics and defense C2 integration protocols [INFERENCE]. Open Source Repositories [KNOWN]
Proprietary Defense C4ISR Platforms Custom defense command systems built by defense contractors. Tailored to military requirements, high security compliance. Extremely costly, long procurement cycles, vendor lock-in, hard to update algorithms quickly [INFERENCE]. Medium High cost and lack of agile, software-only edge retrofit capabilities for legacy IP CCTV installations [PS, INFERENCE]. Defense Tech Publications [RESEARCHED]

9. Candidate Solution Approaches

Approach 1: Centralized Cloud / Sector HQ Processing

  • Mechanism: All remote BOP CCTV streams are compressed and streamed continuously over WAN/VSAT to a centralized cloud/datacenter server running heavy GPU clusters for AI inference.
  • Required Data: High-bandwidth live RTSP video feeds from all cameras simultaneously.
  • Architecture: Edge RTSP Streamer $\rightarrow$ WAN / Satellite Network $\rightarrow$ Central GPU Inference Server (Triton/DeepStream) $\rightarrow$ Web Dashboard.
  • Pros & Cons: High centralized model accuracy and simple edge hardware management. Critical Flaw: Requires continuous 2–4 Mbps per camera WAN bandwidth. In remote border areas with unstable connectivity, the pipeline completely collapses during outages [INFERENCE].
  • Hackathon / Production Feasibility: High for hackathon demo; Completely unviable for production border deployment [INFERENCE].

Approach 2: Proprietary Edge AI Smart Camera Replacement

  • Mechanism: Replace all legacy CCTV cameras with modern edge AI smart cameras featuring built-in NPU chips for on-camera inference.
  • Required Data: On-camera video streams; edge configuration updates.
  • Architecture: AI Camera (Built-in NPU) $\rightarrow$ Local Alert Gateway $\rightarrow$ Central C2.
  • Pros & Cons: Extremely low latency; minimal edge server footprint. Critical Flaw: Directly violates the core requirement of the problem statement ("without requiring dedicated FRS, ANPR, or smart-camera hardware" [PS]) and requires massive capital expenditure.
  • Hackathon / Production Feasibility: Disqualified by problem constraints [PS].

Approach 3: Hybrid Software-Defined Edge-Gateway Architecture (Recommended)

  • Mechanism: Deploy a localized software-defined IBVAP Edge Gateway (e.g., an industrial x86 server or NVIDIA Jetson Orin box) at each BOP. The gateway ingests local legacy camera RTSP streams, executes lightweight quantized AI models (YOLO, ByteTrack, RetinaFace, ANPR OCR) on-site, generates immediate local audio-visual alerts, caches incident clips, and syncs lightweight JSON telemetry (+ compressed snapshot) over low-bandwidth links to the IBVAP Central C2 Platform.
  • Required Data: Local RTSP streams; local edge SQLite/PostgreSQL store; remote MQTT telemetry link.
  • Architecture: Legacy IP Cameras $\rightarrow$ Local LAN $\rightarrow$ IBVAP Edge Gateway (DeepStream/GStreamer + TensorRT) $\rightarrow$ Local BOP UI & Audio Alarm $\rightarrow$ Low-Bandwidth MQTT Store-and-Forward $\rightarrow$ IBVAP Central C2 Master Dashboard.
  • Pros & Cons: Zero modification to existing cameras [PS]; operates 100% autonomously during WAN network blackouts; uses <10 KB/s WAN bandwidth per alert [INFERENCE]; highly scalable across hundreds of BOPs.

Objective Comparison Matrix

Evaluation Criteria Approach 1: Central Cloud Approach 2: AI Hardware Upgrade Approach 3: Hybrid Edge Gateway (Selected)
Problem Statement Fit Low (Bandwidth failure) Disqualified (Violates PS constraint) [PS] 100% Fit [PS, INFERENCE]
Infrastructure Hardware Cost High (Massive WAN/Cloud bandwidth) Extremely High (Total camera replacement) Low (Uses legacy CCTV + cheap edge compute) [PS]
Offline Resilience Zero (Fails when internet drops) Medium High (Full local autonomy & caching) [INFERENCE]
Bandwidth Consumption Very High (~30 Mbps per BOP) Very Low Very Low (<50 KB per alert event) [INFERENCE]
Accuracy Potential High Medium (Constrained on-camera NPU) High (Optimized edge GPU inference) [ASSUMPTION]
Scalability Poor (WAN bottleneck) Moderate High (Decentralized horizontal scale) [INFERENCE]
Hackathon Feasibility High Moderate High (Easily simulated via Docker) [ASSUMPTION]

10. Recommended Approach & Why

The Hybrid Software-Defined Edge-Gateway Architecture (Approach 3) is selected.

Justification & Rejection of Alternatives

  1. Compliance with PS Constraints: Approach 3 strictly fulfills the explicit requirement to leverage existing CCTV infrastructure without hardware modifications to cameras [PS]. Approach 2 was explicitly rejected as it violates this primary constraint.
  2. Bandwidth Realities at Remote Border Outposts: Approach 1 was rejected because streaming continuous raw video feeds over remote border satellite/cellular links is practically impossible and cost-prohibitive [INFERENCE]. Approach 3 processes video locally at the BOP, converting heavy video streams into ultra-lightweight metadata JSON strings and compressed event snapshots for transmission [INFERENCE].
  3. Operational Autonomy: Border outposts frequently experience network cut-offs due to weather or terrain [KNOWN]. Approach 3 guarantees that local BOP guards receive sub-second intrusion alarms and live video overlays even if the external network connection is completely severed [INFERENCE].

11. Technical Architecture

Layered Architecture Diagram

+---------------------------------------------------------------------------------------------------+
|                                 IBVAP SYSTEM ARCHITECTURE                                         |
+---------------------------------------------------------------------------------------------------+

[ LAYER 1: INGESTION LAYER - BOP LOCAL LAN ]
  +----------------------+   +----------------------+   +----------------------+
  | Legacy IP Camera 01  |   | Legacy IP Camera 02  |   | Legacy IP Camera N   |
  | (RTSP/H.264/1080p)   |   | (RTSP/H.265/720p)    |   | (IR Night Vision)    |
  +----------+-----------+   +----------+-----------+   +----------+-----------+
             |                          |                          |
             +--------------------------+--------------------------+
                                        | RTSP Streams (Local LAN)
                                        v
+---------------------------------------------------------------------------------------------------+
| LAYER 2: IBVAP EDGE GATEWAY NODE (DEPLOYED AT BOP)                                                |
|                                                                                                   |
|  +---------------------------------------------------------------------------------------------+  |
|  | Stream Decoding & Pre-processing Engine (GStreamer / MediaMTX / OpenCV)                     |  |
|  | - RTSP Reconnection Buffer  |  Hardware Decoding (NVDEC / VAAPI)  | CLAHE Low-Light Enhancer|  |
|  +----------------------------------------------+----------------------------------------------+  |
|                                                 | Pre-processed Video Frames                   |
|                                                 v                                                 |
|  +---------------------------------------------------------------------------------------------+  |
|  | AI Inference Pipeline (NVIDIA DeepStream / TensorRT Acceleration Engine)                     |  |
|  |                                                                                             |  |
|  |  [Module A: Primary Detector]    [Module B: Tracking]       [Module C: Facial Recognition]   |  |
|  |  YOLOv11 (Person/Vehicle)  --->  ByteTrack (IDs & Motion) --> SCRFD + ArcFace Embeddings     |  |
|  |                                                                                             |  |
|  |  [Module D: ANPR Engine]         [Module E: Virtual Fence]  [Module F: Behavior & Night]    |  |
|  |  Plate Loc + PaddleOCR     --->  Polygon Intersection   --> Dwell-Time & Low-Light Anomaly   |  |
|  +----------------------------------------------+----------------------------------------------+  |
|                                                 | Structured Event Metadata + Bounding Boxes      |
|                                                 v                                                 |
|  +---------------------------------------------------------------------------------------------+  |
|  | Edge Logic Core & Local Storage                                                             |  |
|  | - Rule Engine (Alert Decision & Severity Classifier)                                        |  |
|  | - Local Storage: SQLite / Embedded RocksDB (Event Logs, Crop Snapshots, Video Clips)       |  |
|  | - BOP Local Monitoring Dashboard (WebSockets / WebRTC Live Low-Latency Overlay Feed)        |  |
|  +----------------------------------------------+----------------------------------------------+  |
|                                                 | Offline Store-and-Forward Sync Queue            |
+-------------------------------------------------|-------------------------------------------------+
                                                  | Low-Bandwidth MQTT / NATS (TLS 1.3)
                                                  v
+---------------------------------------------------------------------------------------------------+
| LAYER 3: IBVAP CENTRAL C2 PLATFORM (SECTOR HQ / CENTRAL DATACENTER)                               |
|                                                                                                   |
|  +---------------------------------------------------------------------------------------------+  |
|  | MQTT / Telemetry Message Ingestion Gateway (Mosquitto / NATS JetStream)                      |  |
|  +----------------------------------------------+----------------------------------------------+  |
|                                                 | Sync Payload                                    |
|                                                 v                                                 |
|  +---------------------------------------------------------------------------------------------+  |
|  | Central Microservices Backend (FastAPI / Go Services)                                       |  |
|  | - Watchlist Manager (Distribution of Criminal Faces & Blacklisted Plates to Edges)         |  |
|  | - Multi-BOP Incident Analytics & Cross-Camera Re-ID Query Engine                            |  |
|  | - C2 Integration Gateway (REST API / Webhooks / Cursor-on-Target Encoders)                   |  |
|  +----------------------------------------------+----------------------------------------------+  |
|                                                 | Spatial & Incident Data                         |
|                                                 v                                                 |
|  +---------------------------------------------------------------------------------------------+  |
|  | Central Database & Master Vector Store                                                      |  |
|  | - PostgreSQL + PostGIS (GIS Locations, Incident Master Records, Audit Logs)                   |  |
|  | - Milvus Vector DB (Master Face Embeddings Engine)                                            |  |
|  +----------------------------------------------+----------------------------------------------+  |
|                                                 | Real-Time Map Data & Incident Alerts            |
|                                                 v                                                 |
|  +---------------------------------------------------------------------------------------------+  |
|  | Sector HQ Command GIS Dashboard (Next.js 14 + MapLibre GL)                                   |  |
|  | - Multi-BOP Threat Heatmaps  |  Watchlist Match Alerts  |  C2 System Export Controls        |  |
|  +---------------------------------------------------------------------------------------------+  |
+---------------------------------------------------------------------------------------------------+

Step-by-Step Data Flow Narrative

  1. Stream Ingestion & Decoding: Legacy IP cameras push RTSP H.264/H.265 video feeds over local BOP Ethernet/Wi-Fi to the IBVAP Edge Gateway [PS, ASSUMPTION]. The ingestion engine decodes video frames using hardware acceleration (NVDEC/VAAPI) and maintains reconnect buffers to ensure stream persistence [ASSUMPTION].
  2. Night & Contrast Pre-processing: Low-light or night IR feeds undergo dynamic contrast enhancement (CLAHE filter) to amplify edge details before inference [ASSUMPTION].
  3. Primary Object Detection & Tracking: Frames are passed to a TensorRT-optimized YOLOv11 detector, generating bounding boxes for person and vehicle classes [ASSUMPTION]. Bounding boxes are tracked across frames by ByteTrack to assign consistent track IDs and calculate speed/direction vectors [ASSUMPTION].
  4. Specialized Parallel Analytics:
    • ANPR Pipeline: Vehicle bounding boxes trigger plate region extraction, which is passed to PaddleOCR to extract text strings [RESEARCHED, ASSUMPTION].
    • Facial Biometrics Pipeline: Person bounding boxes containing faces trigger SCRFD face detection and ArcFace feature extraction (512-d vector) [RESEARCHED, ASSUMPTION]. Vectors are matched locally against the cached watchlist using Cosine Similarity [ASSUMPTION].
    • Spatial & Behavioral Analytics: Person/vehicle footprint coordinates are checked against user-defined GeoJSON polygon coordinates (Virtual Fence) using Shapely spatial intersection math [ASSUMPTION]. Dwell times are computed to flag loitering or illegal border crossing [INFERENCE].
  5. Local Alerting & Caching: If an intrusion, watchlist match, or suspicious activity rule triggers, the Edge Core instantly generates a local alert on the BOP Duty Officer's console, playing an audible alarm and displaying a WebRTC video overlay with bounding boxes [INFERENCE]. The event metadata, crop image snapshot, and 10-second video buffer are logged into local SQLite storage [ASSUMPTION].
  6. Low-Bandwidth Store-and-Forward Synchronization: The Edge Gateway packages the alert into a compact JSON payload (<5 KB) with an optional compressed snapshot (<50 KB) and pushes it via MQTT over TLS to Sector HQ [ASSUMPTION]. If the network connection is down, MQTT queues messages locally on disk and auto-flushes them once connectivity is restored [INFERENCE].
  7. Central C2 Visualization & External Integration: Sector HQ receives telemetry events, updating the live MapLibre GIS map with threat markers and logging incidents into PostGIS [ASSUMPTION]. The C2 Integration Gateway exposes REST endpoints and Cursor-on-Target (CoT) XML feeds to forward alerts to legacy MHA command systems [PS, ASSUMPTION].

12. Technology Stack

Layer Selected Technology Why It Fits Alternative Considered Why Rejected
Video Stream Ingestion MediaMTX & GStreamer High-throughput, ultra-low latency RTSP decoding with native C/C++ hardware acceleration binding [ASSUMPTION]. OpenCV VideoCapture Single-threaded, inefficient frame dropping, causes high CPU bottlenecks under multi-camera load [INFERENCE].
Edge AI Framework NVIDIA DeepStream / TensorRT Achieves 10x-20x inference throughput via INT8/FP16 CUDA acceleration on edge GPUs [RESEARCHED]. Plain PyTorch / TensorFlow Extremely slow CPU/GPU inference; cannot maintain real-time 25 FPS across 8+ cameras [INFERENCE].
Object Detection & Tracking YOLOv11 + ByteTrack State-of-the-art accuracy-to-speed balance for person/vehicle detection with robust occlusion handling [RESEARCHED]. DeepSORT / Faster-RCNN DeepSORT requires heavy visual feature extraction per crop; Faster-RCNN is too slow for multi-stream edge [INFERENCE].
ANPR Engine PaddleOCR (Quantized TensorRT) Out-of-the-box support for multi-line alphanumeric extraction fine-tuned for Indian plate fonts (HSRP) [RESEARCHED]. Tesseract OCR Poor accuracy on rotated, low-resolution, or reflective night-time vehicle license plates [RESEARCHED].
Face Recognition Engine SCRFD + ArcFace (InsightFace) High accuracy on low-resolution, angled, and partially occluded faces [RESEARCHED]. OpenCV Haar Cascades + LBPH High false-positive rate, unusable under outdoor uncontrolled lighting conditions [INFERENCE].
Local Vector DB (Edge) FAISS (C++ Python Bindings) Blazing-fast local vector similarity search for <10,000 face watchlists directly in memory [RESEARCHED]. Milvus / Qdrant Cloud Heavy memory footprint; unnecessary overhead for single edge nodes [INFERENCE].
Edge Backend & DB FastAPI + SQLite / Embedded RocksDB Asynchronous, lightweight Python backend with zero-maintenance embedded local database [ASSUMPTION]. Heavy Spring Boot / PostgreSQL on Edge Consumes excessive system RAM and storage resources on edge hardware [INFERENCE].
Central Backend & Master DB FastAPI / Go + PostgreSQL / PostGIS Spatial GIS query capability, high concurrency, robust enterprise-grade reliability [ASSUMPTION]. MongoDB Poor native support for complex spatial GIS polygon queries compared to PostGIS [INFERENCE].
Central Vector Database Milvus Highly scalable, enterprise-grade vector database for million-scale master face watchlists [RESEARCHED]. PGVector Lower throughput performance under high-concurrency vector insertion/search loads [RESEARCHED].
Sync & Messaging Protocol MQTT (Eclipse Mosquitto) Extremely lightweight TCP protocol designed for lossy, low-bandwidth satellite/cellular links [KNOWN]. HTTP REST APIs High header overhead; fails gracefully handling intermittent network drops [INFERENCE].
Central C2 Frontend Next.js 14 + MapLibre GL High-performance vector tile rendering for GIS maps, reactive real-time WebSocket dashboard [ASSUMPTION]. Legacy Angular / Leaflet Slower rendering performance for dense live map markers and multiple video overlays [INFERENCE].

13. AI/ML Justification

Detailed AI Component Matrix

1. Human & Vehicle Detection & Tracking

  • Purpose: Detect and assign persistent IDs to persons and vehicles across camera frames [PS].
  • Model Type: YOLOv11 object detector + ByteTrack tracker [ASSUMPTION].
  • Input: Decoded RGB video frame (1080p/720p at 15–25 FPS) [INFERENCE].
  • Output: Bounding box coordinates [x1, y1, x2, y2], class ID (person, car, truck, bus, two-wheeler), track ID, confidence score [INFERENCE].
  • Data Required: Open-source surveillance datasets (COCO, VisDrone) fine-tuned on custom Indian border terrain images [ASSUMPTION].
  • Why AI: Rule-based background subtraction fails completely due to swaying vegetation, lighting shifts, shadows, and weather changes [INFERENCE].
  • Evaluation Metric: Mean Average Precision ([email protected] > 88%), Higher Order Tracking Accuracy (HOTA > 65%) [ASSUMPTION].
  • Failure Mode: False positives caused by animals (e.g., cattle, dogs) or false negatives during extreme heavy fog [INFERENCE].
  • Human Oversight: Officer verifies alert thumbnail on BOP console before dispatching QRT [ASSUMPTION].

2. Automatic Number Plate Recognition (ANPR)

  • Purpose: Identify vehicle license plate text for blacklist matching [PS].
  • Model Type: Custom License Plate Bounding Box Localizer + PaddleOCR text recognition [RESEARCHED, ASSUMPTION].
  • Input: Cropped vehicle image region [INFERENCE].
  • Output: Alphanumeric license plate string (e.g., UK07AB1234), plate confidence score [INFERENCE].
  • Data Required: Dataset of Indian license plates including High Security Registration Plates (HSRP) and standard formats under diverse angles/lighting [RESEARCHED, ASSUMPTION].
  • Why AI: Traditional pattern matching cannot handle non-standard fonts, dirt, reflection, angled perspectives, and varied state formats [RESEARCHED, INFERENCE].
  • Evaluation Metric: Character Recognition Accuracy (>92%), Plate Localization Recall (>95%) [ASSUMPTION].
  • Failure Mode: Mismatched characters due to muddy plates or customized "fancy" illegible fonts [RESEARCHED, INFERENCE].
  • Human Oversight: OCR output string displayed alongside raw cropped plate snapshot for instant human visual confirmation [ASSUMPTION].

3. Facial Biometric Identification (FRS)

  • Purpose: Detect faces and identify individuals against a POI watchlist [PS].
  • Model Type: SCRFD face detector + ArcFace feature embedder (ResNet50 / MobileFaceNet backbone) [RESEARCHED, ASSUMPTION].
  • Input: Cropped human face region (>60x60 pixels resolution) [INFERENCE].
  • Output: 512-dimensional floating-point vector embedding; matched Person ID + similarity score [INFERENCE].
  • Data Required: Normalized face datasets (LFW, MS1MV2) fine-tuned for low-resolution surveillance crops [RESEARCHED, ASSUMPTION].
  • Why AI: Biometric facial identification is mathematically impossible using traditional deterministic algorithms [INFERENCE].
  • Evaluation Metric: True Accept Rate (TAR) > 95% at False Accept Rate (FAR) = 0.001% [ASSUMPTION].
  • Failure Mode: False dismissal due to severe facial angle (>45 degrees), heavy physical disguise, or extreme darkness [INFERENCE].
  • Human Oversight: High-severity alert triggered only above 0.75 Cosine Similarity; officer visually compares side-by-side face images [ASSUMPTION].

4. Behavioral & Loitering Analytics

  • Purpose: Flag suspicious loitering, sudden running, or movement in restricted directions [PS].
  • Model Type: Heuristic Spatio-Temporal Trajectory Analyzer operating on ByteTrack track histories [ASSUMPTION].
  • Input: Sequence of tracked object centroid coordinates (x, y, t) over time [INFERENCE].
  • Output: Behavioral flag (loitering_detected, rapid_approach, reverse_flow) + anomaly confidence [INFERENCE].
  • Data Required: Trajectory vector logs [INFERENCE].
  • Why AI/Algorithmic Fusion: Tracks spatial-temporal movement patterns across time windows that human operators cannot manually track across 20 feeds simultaneously [INFERENCE].
  • Evaluation Metric: False Alarm Rate < 2 per camera/day [ASSUMPTION].
  • Failure Mode: Nuisance alerts caused by guards on routine patrol standing near boundary lines [INFERENCE].
  • Human Oversight: Configurable dwell-time thresholds (e.g., >30 seconds in zone) set by duty officers [ASSUMPTION].

Where AI Should NOT Be Used

  • Virtual Fence Intersection Math: Determining if an object has crossed a virtual fence boundary must be executed using deterministic computational geometry algorithms (e.g., Ray-Casting Polygon Intersection / Shapely spatial math) on object base coordinates, NOT generative or probabilistic AI models [INFERENCE].
  • User Authentication & Role Management: System access control and audit logging must strictly rely on deterministic cryptographic RBAC mechanisms [INFERENCE].
  • Alert Storage & Sync Rules: Network transmission queuing and store-and-forward retries must rely on deterministic state machines (MQTT QoS 1) [INFERENCE].

14. Data Architecture & Offline/Low-Network Design

Data Schema (Relational & JSON Layout)

-- Local Edge & Central PostGIS Incident Table
CREATE TABLE incident_alerts (
    alert_id UUID PRIMARY KEY,
    bop_id VARCHAR(50) NOT NULL,
    camera_id VARCHAR(50) NOT NULL,
    timestamp TIMESTAMPTZ NOT NULL,
    alert_type VARCHAR(30) CHECK (alert_type IN ('INTRUSION', 'ANPR_MATCH', 'FACE_MATCH', 'LOITERING', 'NIGHT_MOVEMENT')),
    severity VARCHAR(10) CHECK (severity IN ('LOW', 'MEDIUM', 'HIGH', 'CRITICAL')),
    object_class VARCHAR(20),
    track_id INT,
    bounding_box JSONB, -- [x1, y1, x2, y2]
    metadata JSONB,     -- e.g., {"plate_number": "UK07AB1234", "confidence": 0.94}
    snapshot_path VARCHAR(255),
    sync_status VARCHAR(15) DEFAULT 'PENDING' CHECK (sync_status IN ('PENDING', 'SYNCED', 'FAILED')),
    location GEOMETRY(Point, 4323) -- GeoJSON Camera GPS location
);

-- Central Master Watchlist Table
CREATE TABLE face_watchlist (
    person_id UUID PRIMARY KEY,
    full_name VARCHAR(100),
    alias VARCHAR(50),
    threat_category VARCHAR(30),
    face_embedding VECTOR(512), -- Milvus / PGVector extension
    created_at TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP
);

Local Storage & Offline Store-and-Forward Strategy

  • Local Storage Footprint: Each IBVAP Edge Gateway maintains a local circular buffer storage on an industrial NVMe/SSD drive [ASSUMPTION]. Event crop snapshots (50 KB) and 10-second video clips (2 MB) are retained locally for 30 days [ASSUMPTION].
  • Store-and-Forward Queue:
    1. When an alert occurs, the Edge Core writes the alert record to local SQLite storage marked as sync_status = 'PENDING' [ASSUMPTION].
    2. An background MQTT worker attempts to publish the lightweight JSON payload over sih/bop/{bop_id}/alerts with QoS 1 (At Least Once) [ASSUMPTION].
    3. If the external WAN link is disconnected, the payload remains queued in local persistent storage [INFERENCE].
    4. Upon network reconnection, the broker automatically drains the pending queue, transmitting historical alerts chronologically to Central HQ without data loss [INFERENCE].
  • Bandwidth Reduction: Continuous video streaming requires ~3 Mbps per camera. IBVAP's edge telemetry sync requires only ~5 KB per alert JSON metadata, representing a >99.8% bandwidth reduction over satellite/cellular networks [INFERENCE].

15. Security, Privacy & Governance

  1. Biometric & Personal Data Encryption: Face embeddings and license plate logs constitute sensitive operational data [INFERENCE]. All stored biometric vector tables and image snapshots are encrypted at rest using AES-256 [ASSUMPTION]. All edge-to-central MQTT/HTTP communications are forced over TLS 1.3 with mutual certificate authentication (mTLS) [ASSUMPTION].
  2. Access Control & RBAC: Strict Role-Based Access Control governs system functions:
    • BOP Operator: View live feeds, acknowledge local alerts, set temporary virtual tripwires [ASSUMPTION].
    • Sector Analyst: Access historical cross-BOP search, manage watchlists, view macro heatmaps [ASSUMPTION].
    • System Admin: Configure node parameters, review cryptographic audit logs, update AI model weights [ASSUMPTION].
  3. Audit Logging & Non-Repudiation: Every operator action (alert dismissal, watchlist addition, configuration change, video export) is logged to an immutable append-only audit trail [INFERENCE].
  4. Governance & Regulatory Alignment: The system operates under the legal framework governing Central Armed Police Forces (CAPF) along national borders [RESEARCHED]. Data retention policy mandates auto-purging of non-incident video buffer storage after 14 days, retaining only verified alert event logs [ASSUMPTION].

16. MVP Scope (Hackathon Build)

In Scope (MUST BUILD & DEMO)

  • Multi-Stream RTSP Video Ingestion: Concurrent decoding of 3 simulated RTSP IP camera feeds via local MediaMTX/FFmpeg [ASSUMPTION].
  • Real-time Detection & Tracking: YOLOv11 + ByteTrack processing running at >20 FPS on developer laptop GPU [ASSUMPTION].
  • Virtual Fence Intrusion Alerting: Interactive UI drawing of dynamic polygon tripwires on video feeds with sub-second visual/audible alert triggers [ASSUMPTION].
  • ANPR Pipeline: Detection and OCR extraction of Indian High Security Registration Plates (HSRP) from vehicle video clips [RESEARCHED, ASSUMPTION].
  • Facial Recognition Watchlist Matching: Face detection and vector matching against a mock watchlist of 5 target faces with confidence thresholding [ASSUMPTION].
  • Night-time Movement Detection: CLAHE pre-processing filter demonstrating enhanced motion detection on low-light black-and-white IR test video feeds [ASSUMPTION].
  • Offline Store-and-Forward Simulation: Live demonstration of severing the network connection, triggering local alerts, queuing payloads, and auto-syncing to the Central C2 Map upon reconnect [ASSUMPTION].
  • Central C2 GIS Dashboard: Interactive MapLibre map displaying simulated BOP locations, real-time alert markers, incident popups, and watchlist search [ASSUMPTION].

Described Only / Out of Scope (PRODUCTION ONLY)

  • Direct hardware integration with proprietary physical MHA C4ISR defense networks [PS, ASSUMPTION].
  • Million-scale distributed Milvus cluster deployment across physical data centers [ASSUMPTION].
  • Industrial hardware hardening (IP67 enclosure testing, thermal dissipation validation) [ASSUMPTION].
  • Over-The-Air (OTA) firmware/model update delivery pipelines across air-gapped networks [ASSUMPTION].

17. Live Demo Simulation Plan

To ensure a flawless SIH presentation without reliance on unreliable external internet or physical border feeds:

  1. Local RTSP Stream Simulation:

    • Pre-record 3 high-definition test video loops representing:
      • Feed A: Day-time border road with approaching vehicles displaying HSRP license plates [ASSUMPTION].
      • Feed B: Perimeter fence line where a simulated intruder crosses a restricted boundary line [ASSUMPTION].
      • Feed C: Night-time black-and-white IR feed showing a low-light moving figure [ASSUMPTION].
    • Run an instance of MediaMTX (open-source RTSP server) locally on the demo laptop, streaming the pre-recorded video files on loop to local RTSP URLs (rtsp://localhost:8554/cam1, /cam2, /cam3) [ASSUMPTION].

  2. Step-by-Step Demo Flow for SIH Judges:

    • Act I: Real-time Multi-Stream Monitoring & Virtual Fence Setup. Show the duty officer console loading 3 live RTSP streams. Draw a dynamic red polygon boundary (Virtual Fence) across cam2.
    • Act II: Intrusion & Suspicious Behavior Trigger. Allow the test video to progress. As the simulated intruder crosses the drawn polygon boundary, show the instantaneous red flashing bounding box, audible alarm, and alert entry in the event log (<500ms latency) [ASSUMPTION].
    • Act III: ANPR & Watchlist Face Match. Transition to cam1. The system detects a vehicle, crops the plate region, extracts string UK07AB1234, matches it against the local blacklist, and displays a "Blacklisted Vehicle Detected" warning popup. Simultaneously trigger a face match against a POI watchlist [ASSUMPTION].
    • Act IV: Offline Resilience Demonstration. Unplug the laptop Ethernet / disable local network interface. Trigger a new intrusion alert on cam3. Show that the BOP local screen still alarms and saves the incident locally. Re-enable network; show MQTT auto-syncing the cached alert to the Central Sector HQ GIS Map in real-time [ASSUMPTION].

18. Implementation Roadmap

Phase Schedule Leading to 30 September 2026 Deadline

2026 Timeline:
[Mar - Apr] : Phase 1 - Requirements, Architecture & Dataset Curation
[May - Jun] : Phase 2 - Core Engine & Pipeline Development
[Jul - Aug] : Phase 3 - Integration, Frontend UI & GIS Dashboard
[Sep 01-15] : Phase 4 - System Testing, Offline Optimization & Hardening
[Sep 16-30] : Phase 5 - Final Demo Rehearsal & SIH Submission Deployment

Detailed Milestones & Dependencies

  • Phase 1: Foundation & Data Preparation (March 01 – April 30, 2026)

    • Milestone 1.1: Finalize architectural specifications and setup Dockerized development environments.
    • Milestone 1.2: Collect and curate Indian license plate (HSRP) and border surveillance video datasets [RESEARCHED, ASSUMPTION].
    • Milestone 1.3: Annotate custom classes and setup PyTorch training pipelines [ASSUMPTION].

  • Phase 2: Edge AI Pipeline Development (May 01 – June 30, 2026)

    • Milestone 2.1: Build GStreamer/MediaMTX RTSP ingestion engine handling multi-camera frame buffers [ASSUMPTION].
    • Milestone 2.2: Train and export TensorRT INT8 optimized YOLOv11 detection and ByteTrack tracking pipelines [ASSUMPTION].
    • Milestone 2.3: Implement and benchmark PaddleOCR ANPR and ArcFace face embedding extraction [RESEARCHED, ASSUMPTION].
    • Milestone 2.4: Implement CLAHE low-light enhancement pre-processing filter [ASSUMPTION].

  • Phase 3: Logic Core, Messaging & UI Development (July 01 – August 31, 2026)

    • Milestone 3.1: Implement Shapely polygon spatial virtual fencing rule engine [ASSUMPTION].
    • Milestone 3.2: Develop local SQLite store-and-forward queue with MQTT Mosquitto integration [ASSUMPTION].
    • Milestone 3.3: Build BOP Operator Web UI (Next.js/WebSockets) with WebRTC video overlays [ASSUMPTION].
    • Milestone 3.4: Build Central HQ GIS Dashboard (MapLibre GL + PostGIS) [ASSUMPTION].

  • Phase 4: Optimization, Testing & Hardening (September 01 – September 15, 2026)

    • Milestone 4.1: Conduct simulated network disconnection and store-and-forward recovery stress tests [ASSUMPTION].
    • Milestone 4.2: Benchmark FPS, GPU memory usage, and inference latency under 8 concurrent RTSP streams [ASSUMPTION].
    • Milestone 4.3: Fix UI/UX bugs, enforce RBAC security controls, and perform audit log validation [ASSUMPTION].

  • Phase 5: Final Submission & Demo Freeze (September 16 – September 30, 2026)

    • Milestone 5.1: Freeze code repository; deploy self-contained Docker Compose stack [ASSUMPTION].
    • Milestone 5.2: Record system video walkthrough, complete documentation, and finalize SIH dossier [ASSUMPTION].

19. Testing & Validation Plan

Test Category Target Metric / Test Case Success Criteria Validation Methodology
Detection Performance Person/Vehicle Detection Precision & Recall [email protected] $\ge 88%$; Recall $\ge 90%$ Evaluate against 1,000 annotated test frames covering daytime, dusk, and night conditions [ASSUMPTION].
ANPR OCR Accuracy License Plate Extraction Accuracy $\ge 90%$ string accuracy on HSRP plates Test against 300 test vehicle pass clips under varying angles and speeds [RESEARCHED, ASSUMPTION].
Facial Biometrics False Accept Rate (FAR) & True Accept Rate (TAR) TAR $\ge 95%$ at FAR = 0.001% Benchmark vector cosine similarity against 100 face watchlist test gallery [ASSUMPTION].
Inference Latency Edge Frame Processing Throughput $\ge 20$ FPS per stream across 4 concurrent 1080p feeds Measure TensorRT frame loop execution time on target GPU [ASSUMPTION].
Virtual Fence Latency Alert Trigger Latency $< 500$ ms from frame capture to audio-visual alarm Stopwatch timer validation from spatial intrusion frame to WebSocket UI update [ASSUMPTION].
Offline Sync Resilience Network Disconnection Recovery 100% telemetry queue recovery with zero event loss Sever MQTT TCP connection for 15 minutes during active alert generation; re-establish link and verify PostGIS log parity [ASSUMPTION].
False Alarm Testing Rain, Shadow, and Vegetation Interference $< 2$ false intrusion alarms per camera / 24 hours Run 24-hour continuous pre-recorded video loop containing heavy rain, blowing trees, and shadow movements [ASSUMPTION].
Security Validation TLS Encryption & Authentication Zero unencrypted plain-text credentials or streams on local network Wireshark packet capture analysis across edge-to-central transmission channels [ASSUMPTION].

20. Risk Register

Risk Event Likelihood Impact Severity Mitigation Strategy Evidence Tag
Camera Feed Degradation (Heavy Fog / Rain) High High High Apply dynamic contrast enhancement (CLAHE) filters and combine frame-to-frame motion persistence checks before firing intrusion alarms [ASSUMPTION]. [KNOWN]
Edge Hardware GPU Overheating in BOPs Medium High High Enforce automatic frame-skipping logic (e.g., process 10 FPS instead of 25 FPS) when edge GPU junction temperature exceeds $85^\circ\text{C}$ [ASSUMPTION]. [KNOWN]
High False Intrusion Alarms from Animals High Medium Medium Fine-tune YOLO detection classes to strictly separate person and vehicle from animal (cattle, wild dogs); apply minimum bounding box area thresholds [ASSUMPTION]. [INFERENCE]
Network Outage Exceeding Local Storage Capacity Low High Medium Implement FIFO circular storage retention; prioritize retaining text metadata and compressed snapshots while auto-purging raw video buffers first [ASSUMPTION]. [INFERENCE]
Low-Resolution / Blurry License Plates at Night High High High Implement multi-frame OCR consensus (run OCR across 5 consecutive tracking frames and select string with highest confidence score) [ASSUMPTION]. [RESEARCHED]
Team Over-scoping During 36-Hour Hackathon High Medium Medium Enforce strict MVP feature boundary (Section 16); mock external C2 integration endpoints via simple REST webhooks [ASSUMPTION]. [ASSUMPTION]

21. Cost & Scalability

Hackathon Prototype Cost (Zero Budget Setup)

Item Description Cost (INR)
Edge & Central Compute Student personal laptops (NVIDIA RTX 3060/4060 GPU) ₹0 (Existing hardware)
Camera Feed Simulation MediaMTX RTSP Server + FFmpeg open-source tools ₹0 (Free Open Source)
AI Models & Libraries YOLOv11, PyTorch, TensorRT, PaddleOCR, InsightFace, OpenCV ₹0 (Open Source)
Database & Message Broker PostgreSQL, PostGIS, Milvus, SQLite, Mosquitto MQTT ₹0 (Open Source)
GIS Mapping MapLibre GL + OpenStreetMap vector tiles ₹0 (Free Open Source)
Total Hackathon Cost ₹0 (100% Free / Open Source)

Production Deployment Scaling (Per 100 BOP Outposts / ~500 Cameras)

Infrastructure Component Unit Specification Estimated Unit Cost Quantity Required Total Cost
Existing IP CCTV Cameras Legacy RTSP IP Cameras already installed at BOPs [PS] ₹0 (Existing Infra) [PS] 500 Cameras ₹0 [PS]
BOP Edge Gateway Server Industrial Fanless PC / NVIDIA Jetson Orin Industrial (32GB) ~₹1,20,000 per BOP [ASSUMPTION] 100 BOP Nodes ~₹1.20 Crore [ASSUMPTION]
Central C2 Master Server Rackmount GPU Server (Dual NVIDIA L4 / RTX 6000) ~₹8,00,000 per server [ASSUMPTION] 2 Servers (HA Cluster) ~₹16.00 Lakhs [ASSUMPTION]
Software Platform Licensing IBVAP Software Suite (Open-source base + custom code) ₹0 Perpetual License [ASSUMPTION] Enterprise ₹0 [ASSUMPTION]
Installation & Networking Local LAN integration & power backup hardening ~₹15,000 per BOP [ASSUMPTION] 100 BOPs ~₹15.00 Lakhs [ASSUMPTION]
Total Production Estimate Software-Defined Edge Retrofit for 100 BOPs (500 Cameras) ~₹1.51 Crore [ASSUMPTION]

Scalability Justification: Alternative approach of replacing 500 cameras with specialized smart AI cameras and laying dedicated fiber lines would exceed ₹8–10 Crores [INFERENCE]. IBVAP achieves an estimated >80% capital expenditure savings while maintaining complete software modularity [INFERENCE].

22. Differentiation

Parameter Standard Generic Hackathon Submissions Proposed IBVAP Platform
Hardware Dependency Proposes buying expensive new AI cameras or high-cloud streaming [INFERENCE]. 100% Software-Defined: Retrofits existing legacy IP CCTV infrastructure via standard RTSP feeds [PS].
Network Architecture Assumes stable continuous cloud internet connectivity [INFERENCE]. Edge-Hybrid & Offline First: Local inference at BOP with store-and-forward MQTT sync (<5 KB payload) [PS, INFERENCE].
ANPR Capability Uses generic Tesseract OCR failing on Indian plates [RESEARCHED]. Native HSRP Pipeline: Multi-frame PaddleOCR consensus tailored for Indian High Security Registration Plates [RESEARCHED, ASSUMPTION].
C2 Integration Standalone isolated UI without enterprise export capabilities [INFERENCE]. Open Defense C2 Standards: Exposes REST, Webhooks, and Cursor-on-Target (CoT) XML feeds for MHA systems [PS, ASSUMPTION].
Low-Light Handling Fails or ignores night IR video feed degradation [INFERENCE]. Dynamic Night Contrast Pre-processing: Integrated CLAHE filters for standard non-thermal IR cameras [PS, ASSUMPTION].

23. Future Scope

  1. Acoustic & Drone Sensor Fusion: Ingest thermal feeds from aerial drones and acoustic gunshot/intrusion sensors into the IBVAP Edge Gateway for multi-modal threat verification [INFERENCE].
  2. Cross-BOP Vehicle & Person Re-Identification (Re-ID): Implement feature-embedding cross-camera tracking across consecutive BOP outposts along border highways to calculate transit speed and detect detour anomalies [INFERENCE].
  3. Federated Model Learning: Train edge AI models on local edge failure cases at BOPs and push privacy-preserved gradient updates back to Sector HQ without transferring raw video feeds [INFERENCE].

24. Assumptions & Open Items Log

Consolidated Assumptions Log

  • [ASSUMPTION 1] Industrial x86 or NVIDIA Jetson Orin edge nodes are deployable at BOPs with stable local AC/DC generator power.
  • [ASSUMPTION 2] Legacy CCTV IP cameras expose standard unencrypted RTSP video streams over local Ethernet LAN.
  • [ASSUMPTION 3] Local watchlist databases at BOPs contain <10,000 faces, enabling sub-10ms in-memory FAISS similarity search.
  • [ASSUMPTION 4] MQTT protocol over TLS 1.3 is permitted for transmitting low-bandwidth telemetry across government satellite/cellular links.
  • [ASSUMPTION 5] Pre-recorded simulated test video loops accurately represent real-world camera field-of-view, height, and lighting conditions during the SIH demo.

Consolidated Open Items & Validation Needed Log

  • [NEEDS VALIDATION 1] Exact physical RTSP stream resolution, frame rate, and compression codec defaults on deployed SSB cameras across different sectors.
  • [NEEDS VALIDATION 2] Official MHA / SSB C2 software protocol specifications (e.g., specific REST schema or Cursor-on-Target version requirements).
  • [NEEDS VALIDATION 3] Actual low-light performance of existing camera IR illuminators under heavy mountain fog and rain along the Indo-Nepal/Indo-Bhutan borders.
  • [NEEDS VALIDATION 4] Specific government security accreditation and clearance standards required for deploying custom containerized Linux software on CAPF edge networks.

Final Recommendation

Build and demonstrate the IBVAP Software-Defined Hybrid Edge-Gateway Architecture. Do not attempt to stream raw video feeds to the cloud or suggest replacing existing CCTV cameras. Focus the team's build time strictly on executing a rock-solid, local multi-stream RTSP processing engine with instant virtual fencing, ANPR, face matching, night enhancement, and a working offline store-and-forward MQTT sync mechanism.

Recommended Stack

  • Video Ingestion: MediaMTX + GStreamer
  • Edge AI Engine: NVIDIA DeepStream / TensorRT
  • Detection & Tracking: YOLOv11 + ByteTrack
  • ANPR & Biometrics: PaddleOCR + SCRFD / ArcFace
  • Local / Central DB: SQLite (Edge) + PostgreSQL / PostGIS / Milvus (Central)
  • Messaging & Sync: Eclipse Mosquitto MQTT (QoS 1)
  • Frontend Dashboard: Next.js 14 + WebSockets + MapLibre GL

Biggest Dependency

Access to representative video test footage capturing Indian vehicle High Security Registration Plates (HSRP) and outdoor perimeter intrusions under varying lighting conditions.

Biggest Risk

Performance bottlenecks or frame drops on edge hardware when attempting to run concurrent detection, tracking, ANPR, and facial recognition across multiple 1080p camera streams simultaneously.

Biggest Technical Challenge

Engineering an ultra-low latency, multi-threaded C++/Python video pipeline that handles dynamic RTSP stream reconnects while running TensorRT model inference in parallel without memory leaks or buffer overflows.

First 5 Actions

  1. Set up a local MediaMTX RTSP server on the team's development machine and test loop-streaming pre-recorded sample videos.
  2. Build a baseline Python script ingesting the RTSP stream, applying TensorRT-optimized YOLOv11 and ByteTrack, and displaying live FPS metrics.
  3. Integrate Shapely polygon intersection math onto video frame coordinates to establish the Virtual Fence intrusion trigger.
  4. Integrate PaddleOCR on cropped vehicle bounding boxes to test license plate text extraction.
  5. Set up a local Mosquitto MQTT broker and write the offline store-and-forward queue script in SQLite.

Validation Before Coding

  1. Verify that the team's local development machine possesses an NVIDIA GPU with working CUDA, TensorRT, and CUDNN drivers.
  2. Validate that MediaMTX reliably streams test video files over RTSP without dropping frames or introducing synthetic latency (>100ms).
  3. Confirm that sample Indian license plate images can be successfully parsed by PaddleOCR with >90% character accuracy.

Buildability Score /100

92 / 100 (Highly feasible for a skilled 4–6 student team using existing open-source computer vision libraries, standardized web frameworks, and local RTSP video simulation).

Confidence

HIGH (The proposed solution strictly complies with all explicit Ministry of Home Affairs / SSB requirements, respects remote border environment realities, addresses bandwidth constraints, and relies on proven production engineering patterns).

Research Sources

Source Title Source URL Source Type Supported Claim Reliability
Sashastra Seema Bal - Wikipedia Sashastra Seema Bal Wikipedia Reference Article SSB mandate guarding Indo-Nepal (1,751 km) and Indo-Bhutan (699 km) borders under MHA. High
Siliguri: Securing Gateway to India's Northeast Times of India Defence News News / Strategic Analysis Operational terrain, strategic importance of border outposts, and multi-agency security setup. High
Indo-Tibetan & Indo-Nepal Border Security Roles ResearchGate Publication Academic Paper Geography, infrastructure constraints, and technological tools utilized by SSB along open borders. High
Video Analytics Surveillance Integration Guide (2026) Fora Soft Technical Guide Industry Technical Report DeepStream / YOLO / ByteTrack multi-stream edge capabilities, bandwidth math (1.6TB/mo per camera), and edge-cloud hybrid models. High
AI-Powered Video Analytics Security Playbook Fora Soft Security Playbook Industry Technical Report False-alarm reduction metrics (up to 90%), retention pricing, and ONVIF VMS integration practices. High
AI-Powered ANPR Software India Tentosoft Technical Product Page Industry Specification Indian High Security Registration Plate (HSRP) formats, state plate variations, and OCR challenges. High
Analysing HSRP of Vehicles in Indian Scenario IJCRT Research Paper Academic Publication YOLO plate localization accuracy (94%) and OCR extraction pipelines on Indian HSRP plates. High
02 / What our

DEVELOPERS SAY

" 30M+ Google Gemini tokens went into researching, analyzing, and understanding all 233 SIH problem statements to make them easier to explore and solve. "
Satyam Kumar
Satyam Kumar
CEO , Full Stack Developer
" The platform makes SIH problem statements much easier to understand. Having everything from the problem details to the procedure in one place. "
Rishabh Ranjan
Rishabh Ranjan
COO , Content Writer
" Oh, you got ads? Well, we need to make money too! But, you can use our platform to help you win SIH and take home the prize money. "
Himanshu Kumar
Himanshu Kumar
CMO , Product Reviewer
03 / Are You

READY TO
BUILD?

Get in touch if you need help bringing your ideas to life, or simply have an interesting idea you'd like to talk through. I'm always open to a conversation and exploring new possibilities together.

LinkedIn
[email protected]