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.
Search for your problem statement, and start building your solution.
| Description: | • Background Urban public transport buses traverse almost every major road in a city every day. Modern buses are increasingly equipped with multiple cameras covering the front, rear, sides, and passenger cabin. However, these cameras are primarily used for recording incidents and are not leveraged as intelligent sensing platforms. At the same time, city authorities rely on fixed CCTV cameras, manual inspections and citizen complaints to identify road defects, traffic congestion,missing infrastructure and unsafe driving behaviour. This results in delayed response,incomplete situational awareness and inefficient maintenance planning. • Description Develop an AI-powered onboard and centralized software platform that transforms public transport buses into mobile urban sensing units. The onboard software shall analyse video streams from multiple bus-mounted cameras to detect road defects such as potholes, damaged roads, missing road dividers, missing zebra crossings, damaged or missing traffic signboards,waterlogging and other road hazards. It shall estimate vehicle density through vehicle detection, classification and counting, identify traffic bottlenecks, and detect vulnerable pedestrian situations such as school children crossing roads. During incidents such as hit-and-run or rash driving, the system should detect and track the offending vehicle, extract the registration number with a confidence score, timestamp and GPS location, and securely share alerts with a central command system. The centralized platform shall aggregate information from the entire bus fleet, visualize events on a GIS map, generate congestion heat maps,identify infrastructure deficiencies, analyse origin–destination traffic patterns, estimate route delays and provide actionable insights for transport authorities. • Expected Solution The solution should provide an edge-AI onboard processing framework integrated with a centralized urban intelligence platform. It should generate reliable alerts, GIS-based dashboards, road condition maps, traffic analytics and incident reports to support proactive road maintenance, improved traffic management, enhanced public safety and evidence-based decision making while minimizing bandwidth through intelligent edge processing. |
| Organization: | Bharat Electronics Limited |
| Department: | Bharat Electronics Limited |
| Category: | Software |
| Theme: | Smart Automationtd> |
| Youtube Link: | Not provided |
| Dataset Link: | Not provided |
| Contact info: | Not provided |
Public transport buses traverse almost every major urban roadway daily, offering an ideal mobile platform for city-wide sensing [PS]. Modern buses carry cameras covering the front, rear, sides, and passenger cabin, but these cameras currently act as passive incident recorders rather than intelligent sensing networks [PS]. Meanwhile, municipal authorities rely on static CCTV, periodic manual road inspections, and reactive citizen complaints, leading to delayed repairs, incomplete situational awareness, and inefficient traffic management [PS]. This dossier presents an AI-Powered Mobile Urban Intelligence Platform that equips public transit fleets with edge-AI inference units to analyze multi-camera video streams in real time [PS, INFERENCE]. The onboard edge unit detects road defects (potholes, damaged pavement, missing dividers, missing zebra crossings, damaged traffic signs, waterlogging), measures traffic density, identifies bottlenecks, detects vulnerable pedestrians, and tracks hit-and-run or rash driving incidents with Automatic Number Plate Recognition (ANPR), timestamping, and GPS tagging [PS]. Critical alerts and compressed metadata are transmitted via low-bandwidth cellular MQTT to a centralized Geographic Information System (GIS) command dashboard deployed within Bharat Electronics Limited’s (BEL) Smart City Integrated Command and Control Centre (ICCC) architecture [PS, RESEARCHED].
| Parameter | Value |
|---|---|
| Problem Statement Number | SIH26124 [PS] |
| Title | AI-Powered Mobile Urban Intelligence Platform Using Public Transport Fleet [PS] |
| Organization | Bharat Electronics Limited (BEL) [PS] |
| Department | Bharat Electronics Limited (BEL) [PS] |
| Category | Software [PS] |
| Theme | Smart Automation [PS] |
| Submission Deadline | 30 September 2026 [PS] |
| Ideas Submitted Counter | 0/500 [PS] |
| Target Deployment Platform | Onboard Edge Processing Unit (NVIDIA Jetson / Edge AI Box) + Centralized GIS Command Portal [PS, INFERENCE] |
To transform routine public transport bus fleets into real-time, low-bandwidth mobile urban sensing units that continuously inspect road infrastructure, analyze traffic density and origin-destination patterns, detect safety incidents with ANPR, and deliver actionable GIS-mapped insights to municipal command centers [PS, INFERENCE].
The problem description from Bharat Electronics Limited (BEL) contains the following explicit requirements [PS]:
| ID | Source Text / Requirement | Interpretation | Technical & Implementation Implication | Evidence Tag |
|---|---|---|---|---|
| a | "The onboard software shall analyse video streams from multiple bus-mounted cameras to detect road defects such as potholes, damaged roads, missing road dividers, missing zebra crossings, damaged or missing traffic signboards, waterlogging and other road hazards." | Multi-camera onboard Computer Vision model must detect 7 distinct categories of physical road defects and infrastructure hazards in real-time. | Deploy multi-class object detection model (e.g., YOLOv11) on edge GPU. Camera inputs must cover front and side road views. Requires class-wise confidence filtering and geo-referencing using sync GPS. | [PS] |
| b | "It shall estimate vehicle density through vehicle detection, classification and counting, identify traffic bottlenecks, and detect vulnerable pedestrian situations such as school children crossing roads." | Edge software must count and classify vehicles (cars, two-wheelers, auto-rickshaws, buses, trucks), calculate density/flow rate, tag traffic gridlock, and detect pedestrians/children in high-risk zones. | Multi-object tracking (MOT) pipeline (e.g., ByteTrack/DeepSORT) with ROI counting zones. Pedestrian detector fine-tuned for small-stature individuals (children) and proximity to moving vehicles or un-marked crossings. | [PS] |
| c | "During incidents such as hit-and-run or rash driving, the system should detect and track the offending vehicle, extract the registration number with a confidence score, timestamp and GPS location, and securely share alerts with a central command system." | System must trigger incident detection for anomalous vehicle motion (rash driving, collision), track offending vehicle, perform ANPR/OCR, attach GPS + timestamp, and dispatch immediate secure alert. | Edge buffer retains 15-second pre-event and post-event video. Motion trajectory analysis triggers ANPR module (yolov8-plate + PaddleOCR/LPRNet) yielding plate string, confidence %, ISO timestamp, and GPS coords pushed over cellular MQTT. | [PS] |
| d | "The centralized platform shall aggregate information from the entire bus fleet, visualize events on a GIS map, generate congestion heat maps, identify infrastructure deficiencies, analyse origin–destination traffic patterns, estimate route delays and provide actionable insights for transport authorities." | Central web platform aggregates multi-bus edge telemetry, provides GIS mapping, builds heatmap layers, catalogs road defects, computes OD matrices from bus/traffic movement, and forecasts delays. | Web GIS architecture (PostGIS, GeoServer/Mapbox, Leaflet). Spatial indexing (H3 / S2 grid), clustering algorithms for pothole deduplication, and trajectory processing for origin-destination (OD) estimation. | [PS] |
| e | "The solution should provide an edge-AI onboard processing framework integrated with a centralized urban intelligence platform... minimizing bandwidth through intelligent edge processing." | Heavy video processing happens locally on bus hardware; only lightweight structured JSON metadata, alerts, and keyframes are transmitted over LTE/4G, keeping data transfer costs minimal. | Zero continuous cloud video streaming. Local edge inference; MQTT publish payload < 5 KB per event; bulk video upload occurs only via depot Wi-Fi or explicit command request. | [PS] |
| f | "It should generate reliable alerts, GIS-based dashboards, road condition maps, traffic analytics and incident reports to support proactive road maintenance, improved traffic management, enhanced public safety and evidence-based decision making." | Reporting system must export work orders for municipal road departments, traffic police incident logs, and transit delay dashboards. | RBAC role-based dashboards: Municipal Public Works Dept (PWD), Traffic Police, City Bus Operator (RTC/ISTLS). Automated PDF/CSV work-order generator. | [PS] |
| Stakeholder | Goal | Current Pain | System Need |
|---|---|---|---|
| Municipal Road & PWD Engineers | Timely identification and repair of potholes, missing dividers, broken signs, and waterlogging. | Rely on manual field surveys and citizen complaints on portals; delayed repairs lead to accidents and litigation [PS, KNOWN]. | Automated, GIS-mapped road deficiency catalog with deduplicated defect clusters, severity index, and automated work orders [PS, INFERENCE]. |
| Traffic Police & Law Enforcement | Fast identification of hit-and-run offenders, rash drivers, and urban congestion bottlenecks. | Static CCTV cameras have blind spots; manual tracking of hit-and-run vehicles takes days or weeks [PS, KNOWN]. | Real-time high-priority alerts with vehicle plate numbers, confidence score, snapshot, and exact GPS coordinates from passing buses [PS, INFERENCE]. |
| Public Transit Authorities (RTC / City Bus Ops) | Monitor fleet performance, understand route delays, optimize bus schedules, and protect pedestrian safety. | Lack of visibility into real-time road conditions causing transit delays; high accident liability near school zones [PS, INFERENCE]. | Route delay analytics, origin-destination traffic pattern visualizer, and pedestrian proximity safety warnings [PS, INFERENCE]. |
| Smart City ICCC Operators (BEL / Municipal Command) | Single pane of glass for urban sensing without adding expensive static CCTV camera infrastructure across thousands of kms. | High capital and maintenance expenditure for fixed roadside sensor nodes [KNOWN]. | Unified web GIS dashboard integrating fleet-wide edge telemetry into existing BEL Smart City ICCC platforms [PS, RESEARCHED]. |
| Bus Drivers & Onboard Operations | Minimal interaction required; zero distraction while driving. | Onboard systems requiring manual input cause safety hazards and driver resistance [INFERENCE]. | Passive background operation on edge device with automatic power management tied to vehicle ignition [INFERENCE]. |
| Proposed Feature / Component | Justifying Requirement | Implementation Details |
|---|---|---|
| Edge-AI Computer Vision Detection Engine | Req a, e [PS] | TensorRT-accelerated YOLO model running on onboard Edge GPU analyzing multi-camera streams for 7 defect classes [PS, INFERENCE]. |
| Multi-Class Vehicle Detector & Counter (MOT) | Req b [PS] | ByteTrack tracking engine on front/side streams counting vehicles by class (2-wheeler, car, bus, truck, auto) to compute lane occupancy & density [PS, INFERENCE]. |
| School Children & Vulnerable Pedestrian Detection | Req b [PS] | Human detection model with spatial bounding-box association near road boundaries and unmarked pedestrian crossings [PS, INFERENCE]. |
| Incident Trigger & Edge ANPR Module | Req c [PS] | Anomaly detector tracking sudden acceleration/swerving or collision impact; triggers high-res frame capture, license plate extraction (PaddleOCR), confidence computation, and MQTT alert [PS, INFERENCE]. |
| Bandwidth-Optimized MQTT Telemetry Sync | Req e [PS] | Transmits JSON payloads (< 5 KB) containing Lat/Long, Timestamp, Defect/Event Type, and Confidence Score over 4G/LTE [PS, INFERENCE]. |
| Central Web GIS Interactive Dashboard | Req d, f [PS] | Leaflet/Mapbox-based GIS portal displaying bus fleet live locations, defect markers, incident pins, and layer toggles [PS, INFERENCE]. |
| Traffic Congestion Heatmap & OD Matrix Builder | Req d [PS] | Spatial aggregation of bus speeds and surrounding vehicle density into H3 hex-grid heatmaps; spatial trajectory extraction for OD flows [PS, INFERENCE]. |
| Road Maintenance Work Order Generator | Req f [PS] | Automated export module producing municipal work-order tickets (PDF/GeoJSON) with defect snapshot, address, and severity rating [PS, INFERENCE]. |
| Environmental Dimension | Reality / Constraint | Technical Adaptation Strategy |
|---|---|---|
| Connectivity | Variable cellular coverage (4G/3G dead zones in urban underpasses, dense corridors, or suburban routes) [KNOWN]. | Offline Store-and-Forward: Local SQLite/LevelDB database buffers event payloads and telemetry. Pushes via MQTT upon reconnection; bulk high-res logs sync over Wi-Fi at bus depot [INFERENCE]. |
| Vehicle Vibration & Power | Heavy chassis vibration, power spikes/voltage drops during bus startup/idle (12V/24V DC bus line) [KNOWN]. | Industrial edge enclosure with shock mounts; wide-input isolated DC-DC power supply (9V–36V) with supercapacitor shutdown protection [INFERENCE]. |
| Camera Hardware & Mounting | Multi-camera setups (front-facing windshield, side-facing, rear-facing) with varying focal lengths, optical dirt, and rain glare [PS, KNOWN]. | Software image enhancement (CLAHE / rain-wipe detection); robust calibration pipelines for varying camera mounting angles across bus models [INFERENCE]. |
| Lighting & Weather | Daytime glare, nighttime low light, rain droplets on windshields, fog, and shadows from elevated structures [KNOWN]. | Multi-spectral training dataset augmentations (darkness, glare, water drops) and brightness-adaptive model thresholding [INFERENCE]. |
| Edge Hardware Limits | Onboard power budget (< 30W) and thermal limits inside non-air-conditioned urban buses [KNOWN]. | Fanless industrial edge box (e.g., NVIDIA Jetson Orin Nano 8GB); FP16/INT8 quantized TensorRT models running at 15–20 FPS per stream [INFERENCE]. |
| Language & Digital Literacy | Municipal road workers and transit supervisors require simple, local-language interface elements [KNOWN]. | Bilingual UI (English + Hindi/Regional State Language); color-coded severity maps (Red/Yellow/Green) and one-click PDF exports [INFERENCE]. |
| Data Type | Purpose | Ideal Real-World Source | Hackathon Substitute | Access Method | Availability | Privacy Sensitivity | Confidence |
|---|---|---|---|---|---|---|---|
| Road Defect & Hazard Video Data | Training & evaluating computer vision model for potholes, cracks, missing markings [PS]. | Municipal road inspection feeds, Indian Driving Dataset (IDD) [RESEARCHED]. | Public datasets: IDD-Detection, Kaggle Indian Road Potholes, Roboflow Road Defects [RESEARCHED]. | Download via Kaggle / GitHub API [RESEARCHED]. | Public Open Access [RESEARCHED]. | None (Public roads) [KNOWN]. | [RESEARCHED] |
| Indian License Plate Images | Training ANPR model for Indian registration formats [PS]. | Traffic enforcement camera repositories [KNOWN]. | DataCluster Indian License Plates Dataset, Kaggle Indian Vehicle Number Plates [RESEARCHED]. | Direct download from Kaggle / Roboflow Universe [RESEARCHED]. | Public Open Access [RESEARCHED]. | Low (Public vehicle plates) [KNOWN]. | [RESEARCHED] |
| Bus Fleet GPS Trajectories | Testing fleet tracking, GIS mapping, OD matrices, route delay calculations [PS]. | Open Transit Data (e.g., Delhi DTC GTFS-RT, Bengaluru BMTC API) [KNOWN]. | Synthetic GPS generator script reproducing city bus routes with simulated delays and speed variations [ASSUMPTION]. | Python script outputting GeoJSON / NMEA streams [ASSUMPTION]. | Fully Available (Generated) [ASSUMPTION]. | None [KNOWN]. | [INFERENCE] |
| City Road Network Shapefiles | Snapping bus coordinates to road segments (Map Matching) [PS]. | OpenStreetMap (OSM) via Overpass API [KNOWN]. | OSM city export (.osm / GeoJSON) for target hackathon city (e.g., New Delhi / Chandigarh / Bengaluru) [KNOWN]. | Overpass API / QGIS export [KNOWN]. | Public Open Access [KNOWN]. | None [KNOWN]. | [KNOWN] |
| Vehicle Incident Video Clips | Demonstrating hit-and-run tracking and rash driving ANPR alert flow [PS]. | Dashcam incident archives [KNOWN]. | Staged dashcam recordings / public YouTube dashcam evaluation clips [ASSUMPTION]. | Local file ingestion into video pipeline [ASSUMPTION]. | Available [ASSUMPTION]. | High (License plate / driver privacy - anonymized for demo) [INFERENCE]. | [INFERENCE] |
| Existing System | Description | Strengths | Weaknesses | Relevance | Remaining Gap Addressed by Proposed System | Source | Confidence |
|---|---|---|---|---|---|---|---|
| Static Smart City ITMS / CCTV (BEL / Honeywell) | Fixed cameras deployed at major city intersections for traffic monitoring and ANPR [RESEARCHED]. | High-resolution stationary capture, direct integration with city ICCC command centers [RESEARCHED]. | High Blind Spot Coverage: Covers < 5% of total city road network; extremely expensive per linear km [INFERENCE]. | Primary competitor/baseline system designed by BEL [RESEARCHED]. | Mobile Sensing Coverage: Extends monitoring to 100% of routes covered by public buses, inspecting inner roads and mid-block deficiencies [PS, INFERENCE]. | BEL Smart City Portal [1.1, 1.2] | [RESEARCHED] |
| Crowdsourced Citizen Portals (e.g., Swachhata / Local Civic Apps) | Citizens manually report potholes and road issues via mobile apps with geo-tagged photos [KNOWN]. | Zero hardware cost on transit vehicles; direct citizen involvement [KNOWN]. | Unreliable & Reactive: Low reporting rate, duplicate reports, inaccurate location tags, non-standard photo quality [INFERENCE]. | Alternative road defect tracking channel [KNOWN]. | Automated & Proactive: Continuous automated inspection without manual intervention; standardized severity metrics and exact spatial deduplication [PS, INFERENCE]. | Public Municipal Portals [KNOWN] | [KNOWN] |
| Commercial AI Dashcam Systems (e.g., CITYROVER / Samsara) | Standalone dashcams mounted on municipal utility vehicles for pothole logging [RESEARCHED]. | Commercialized pothole logging with GIS map dashboard [RESEARCHED]. | Proprietary closed hardware; lacks multi-camera vehicle density analytics, OD pattern estimation, and hit-and-run ANPR alerting tailored to Indian transit buses [INFERENCE]. | Direct global product equivalent [RESEARCHED]. | Integrated Urban Sensing: Combines 7 road defect classes + traffic counting + school zone safety + hit-and-run ANPR + OD matrix analysis in an edge-optimized framework built for BEL ICCC [PS, INFERENCE]. | CITYROVER Documentation [4.3] | [RESEARCHED] |
| Dimension | Approach 1: Full-Cloud Streaming Architecture | Approach 2: Pure Offline Depot Store-and-Forward | Approach 3: Hybrid Edge-AI & Event Telemetry (Recommended) |
|---|---|---|---|
| Mechanism | Onboard cameras stream raw video (RTSP over 4G/5G) continuously to central cloud for GPU inference [INFERENCE]. | Edge device runs full AI inference onboard; logs all events to local disk; syncs data only when bus docks at depot Wi-Fi [INFERENCE]. | Edge device runs full AI inference onboard; publishes lightweight JSON alerts/counts via 4G/MQTT in real time; syncs video logs at depot Wi-Fi [INFERENCE]. |
| Problem Fit | High latency; prohibitive cellular data bandwidth cost (gigabytes/hour/bus) [INFERENCE]. | Zero real-time alert capability; delayed incident response (hit-and-run alerts delivered hours later) [INFERENCE]. | Optimal: Real-time critical alerts (hit-and-run, safety hazards) + low cellular cost + zero stream congestion [PS, INFERENCE]. |
| Accuracy Potential | High (unconstrained cloud GPU resources) [INFERENCE]. | High (full resolution processing locally) [INFERENCE]. | High (FP16/INT8 TensorRT optimized models on Edge GPU) [INFERENCE]. |
| Data Requirements | High network throughput required [INFERENCE]. | High local storage required (128GB - 256GB NVMe) [INFERENCE]. | Moderate local storage + low network bandwidth (< 1 MB/hour cellular) [INFERENCE]. |
| Complexity | High backend video streaming server complexity [INFERENCE]. | High depot Wi-Fi sync orchestration complexity [INFERENCE]. | Moderate balanced complexity across edge and central GIS platform [INFERENCE]. |
| Explainability | Centralized audit logs [INFERENCE]. | Offline batch reports [INFERENCE]. | Real-time event triggers with attached keyframes & confidence scores [INFERENCE]. |
| Cost | Extremely High (Cellular data charges: ~₹15,000/bus/month) [INFERENCE]. | Low operational network cost [INFERENCE]. | Lowest Total Cost: Minimal cellular payload + standard edge box hardware amortized over 5 years [INFERENCE]. |
| Scalability | Poor (Cloud bandwidth & server ingress choke with 1,000+ buses) [INFERENCE]. | High edge scalability, low network dependency [INFERENCE]. | Linear Scalability: Scale to 5,000+ buses without backend performance degradation [INFERENCE]. |
| Hackathon Feasibility | Hard to demonstrate without active high-bandwidth SIMs [INFERENCE]. | Easy to demo offline, but lacks live alert punchiness [INFERENCE]. | Exceptional: Live interactive demonstration with MQTT alerts, edge dashboard, and central GIS maps [INFERENCE]. |
| Production Feasibility | Impractical for municipal budgets [INFERENCE]. | Partially feasible, but fails emergency response SLAs [INFERENCE]. | Production Ready: Aligns perfectly with BEL Smart City and ICCC deployment specs [PS, RESEARCHED]. |
Approach 3: Hybrid Edge-AI Framework with Real-Time Event Telemetry & Depot Sync [PS, INFERENCE].
Approach 3 executes all heavy computer vision processing (defect detection, traffic counting, school zone monitoring, ANPR) on the onboard edge processor [PS, INFERENCE]. When a defect or incident is detected, it package-encodes a micro-payload (JSON containing event type, GPS coordinates, timestamp, confidence score, and compressed 50 KB keyframe) and transmits it instantaneously over MQTT via 4G [INFERENCE]. Heavy video recordings are stored in a rolling 48-hour onboard NVMe buffer and automatically dumped to city depot servers over high-speed Wi-Fi 6 when docking, providing the perfect balance between instant situational awareness and low operational expense [INFERENCE].
+-----------------------------------------------------------------------------------+
| PUBLIC TRANSPORT BUS (EDGE) |
| |
| +--------------------+ +--------------------+ +--------------------+ |
| | Front Road Camera | | Side Road Camera | | Rear/Cabin Camera | |
| +---------+----------+ +---------+----------+ +---------+----------+ |
| | | | |
| +------------------------+------------------------+ |
| | RTSP / GStreamer Video Streams |
| v |
| +-----------------------------------------------------------------------------+ |
| | ONBOARD EDGE AI ENGINE (NVIDIA Jetson) | |
| | | |
| | +--------------------------+ +--------------------------+ | |
| | | TensorRT Video Pipeline | | GPS / NMEA Telemetry Sync| | |
| | +------------+-------------+ +------------+-------------+ | |
| | | | | |
| | v v | |
| | +-------------------------------------------------------+ | |
| | | AI Model Suite (YOLOv11 Defect & Vehicle Detector, | | |
| | | ByteTrack MOT, School Zone Rules, ANPR / OCR Engine) | | |
| | +---------------------------+---------------------------+ | |
| | | | |
| | v | |
| | +-------------------------------------------------------+ | |
| | | Event Rule Engine, Anomaly Detector & Anonymizer | | |
| | +---------------------------+---------------------------+ | |
| | | | |
| | +--------------+--------------+ | |
| | | | | |
| | v v | |
| | +-----------------------+ +-----------------------+ | |
| | | Local NVMe Video Log | | Offline SQLite Buffer | | |
| | | (Rolling 48-Hour) | | (Unsent MQTT Packets) | | |
| | +-----------+-----------+ +-----------+-----------+ | |
+------------------|-----------------------------|----------------------------------+
| |
Depot Wi-Fi | Bulk High-Res Sync | Real-Time Cellular 4G / MQTT
(600 Mbps) | | (< 5 KB JSON Payload)
v v
+-----------------------------------------------------------------------------------+
| BEL CENTRAL URBAN INTELLIGENCE PLATFORM |
| |
| +-----------------------------------------------------------------------------+ |
| | INGESTION & BROKER LAYER | |
| | EMQX Enterprise MQTT Broker / NGINX API Gateway / Geo-TLS Endpoint | |
| +--------------------------------------+--------------------------------------+ |
| | |
| v |
| +-----------------------------------------------------------------------------+ |
| | BACKEND PROCESSING & ANALYTICS ENGINE | |
| | Node.js / Python FastAPI Services | |
| | - Pothole Spatial Clustering & Deduplication Engine (H3 Hex Grid) | |
| | - Traffic Congestion & Delay Estimator Engine | |
| | - Origin-Destination (OD) Trajectory Matrix Calculator | |
| | - Hit-and-Run Alert Dispatcher & Notification Handler | |
| +--------------------------------------+--------------------------------------+ |
| | |
| v |
| +-----------------------------------------------------------------------------+ |
| | DATA STORAGE LAYER | |
| | PostgreSQL + PostGIS (Spatial Database) | Redis (Real-time Fleet Caching) | |
| | MinIO / S3 Object Storage (Event Keyframes & Wi-Fi Sync Video Archives) | |
| +--------------------------------------+--------------------------------------+ |
| | |
| v |
| +-----------------------------------------------------------------------------+ |
| | CENTRALIZED GIS COMMAND DASHBOARD (ICCC FRONTEND) | |
| | React.js + Leaflet/Mapbox GL JS + Tailwind CSS + Chart.js | |
| | - Interactive City Road Map (Defects, Live Bus GPS, Traffic Density) | |
| | - Congestion Heatmaps & Route Delay Visualizer | |
| | - High-Priority Incident Feed (Hit-and-Run ANPR, School Zone Safety) | |
| | - Municipal Work Order Generator & PWD Export Module | |
| +-----------------------------------------------------------------------------+ |
+-----------------------------------------------------------------------------------+
| Layer | Selected Technology | Why It Fits | Alternative Considered | Why Rejected |
|---|---|---|---|---|
| Edge Hardware | NVIDIA Jetson Orin Nano (8GB) / Jetson AGX [INFERENCE] | Hardware Tensor Cores, low power draw (<15W), handles 4x 1080p deep learning streams at real-time speeds [INFERENCE]. | Raspberry Pi 5 + Coral TPU [INFERENCE] | Lacks GPU memory bandwidth for parallel multi-camera deep learning pipelines [INFERENCE]. |
| Edge Computer Vision Framework | NVIDIA DeepStream SDK / TensorRT + OpenCV [INFERENCE] | Hardware-accelerated video decoding (NVDEC) and zero-copy GPU memory pipelines [INFERENCE]. | Pure CPU OpenCV + PyTorch [INFERENCE] | High CPU bottleneck, drops frames above 10 FPS on multi-camera streams [INFERENCE]. |
| Primary AI Detection Model | YOLOv11 / YOLOv8 (Quantized INT8/FP16) [RESEARCHED, INFERENCE] | State-of-the-art speed/accuracy tradeoff for real-time edge detection [RESEARCHED]. | Faster R-CNN / SSD [INFERENCE] | Slower inference speed (> 100ms/frame), unsuited for high-speed bus sensing [INFERENCE]. |
| ANPR / OCR Engine | YOLOv8-Plate + PaddleOCR / LPRNet [RESEARCHED, INFERENCE] | Lightweight, handles multi-font Indian registration plates and dirty/angled surfaces [RESEARCHED]. | Tesseract OCR [INFERENCE] | Fails under non-standard Indian plate fonts, high noise sensitivity [INFERENCE]. |
| Ingestion Protocol | MQTT (EMQX Broker) over TLS [INFERENCE] | Micro-header overhead, ultra-low bandwidth, pub/sub model ideal for intermittent cellular connectivity [INFERENCE]. | HTTP REST APIs [INFERENCE] | Heavy HTTP header overhead, poor handling of intermittent mobile network drops [INFERENCE]. |
| Backend Framework | Python FastAPI + Node.js (Express) [INFERENCE] | Asynchronous I/O, fast JSON processing, native integration with Geospatial and ML libraries [INFERENCE]. | Java Spring Boot [INFERENCE] | Slower development iteration cycle for hackathon prototyping [INFERENCE]. |
| Database & Spatial Indexing | PostgreSQL + PostGIS Extension + Uber H3 Grid [INFERENCE] | Gold standard for enterprise GIS, spatial queries (ST_DWithin), and hexagonal spatial aggregation [INFERENCE]. | MongoDB [INFERENCE] | Weaker spatial join performance and lack of advanced spatial analytics primitives [INFERENCE]. |
| Frontend GIS Dashboard | React.js + Leaflet.js / Mapbox GL JS + Tailwind CSS [INFERENCE] | High-performance vector tile rendering, smooth map pan/zoom with thousands of markers [INFERENCE]. | Google Maps JS API [INFERENCE] | Proprietary licensing costs, rate limits, and restricted offline/custom GIS server integrations [INFERENCE]. |
| AI Component | Purpose | Model Type | Input | Output | Data Required | Why AI is Preferable over Rules | Evaluation Metric | Failure Mode | Human Oversight |
|---|---|---|---|---|---|---|---|---|---|
| Road Defect & Hazard Detector | Detect potholes, damaged roads, missing dividers/crossings, bad signs, waterlogging [PS]. | YOLOv11s Object Detector (Quantized TensorRT) [RESEARCHED, INFERENCE]. | 1080p front/side camera video frames [PS]. | Bounding boxes, class labels, confidence scores [INFERENCE]. | Indian Driving Dataset (IDD), pothole datasets (5,000+ images) [RESEARCHED]. | Hand-crafted edge filters fail under variable asphalt color, shadows, rain reflections, and lighting changes [INFERENCE]. | mAP@50-95 (> 0.82), Inference FPS (> 25) [INFERENCE]. | False positive on shadow/asphalt patch [INFERENCE]. | Spatial deduplication across multiple bus passes; operator verification before PWD work-order release [INFERENCE]. |
| Vehicle Density & Classifier | Classify and count vehicles (car, 2-wheeler, bus, truck, auto) [PS]. | YOLOv11n + ByteTrack MOT engine [INFERENCE]. | Front/rear camera video streams [PS]. | Track IDs, vehicle class, line-crossing count, density score [INFERENCE]. | Traffic vision datasets (DriveIndia, COCO vehicle classes) [RESEARCHED]. | Rule-based motion sensors cannot classify vehicle types or handle occluded urban traffic [INFERENCE]. | Multiple Object Tracking Accuracy (MOTA > 75%) [INFERENCE]. | Missed detection during extreme vehicle tailgating [INFERENCE]. | Aggregated over 15-minute windows; outlier filtering algorithms [INFERENCE]. |
| Vulnerable Pedestrian Safety Model | Flag pedestrians and children near roadways / school zones [PS]. | YOLOv11 Human/Pose Detector [INFERENCE]. | Front camera video stream [PS]. | Pedestrian bounding boxes, estimated height/stature tag, proximity alert [INFERENCE]. | COCO Person, Cityscapes, custom pedestrian datasets [RESEARCHED]. | Heuristic motion rules flag non-human objects; deep learning reliably identifies human morphology [INFERENCE]. | [email protected] (> 0.90 for pedestrian class) [INFERENCE]. | Occlusion behind parked vehicles [INFERENCE]. | Highlights safety warning zone on bus driver dashboard/central log [INFERENCE]. |
| ANPR / License Plate OCR | Extract vehicle plate text during hit-and-run / rash driving incidents [PS]. | Custom YOLOv8 Plate Detector + PaddleOCR [RESEARCHED, INFERENCE]. | Cropped vehicle region bounding box [INFERENCE]. | Plate text string, confidence score % [PS]. | Indian Vehicle License Plate Dataset (10,000+ annotated plates) [RESEARCHED]. | Template matching fails due to multi-line fonts, non-standard state codes, and dirt/dust [RESEARCHED]. | Character Accuracy (> 92%), Exact Match Precision (> 85%) [INFERENCE]. | Misreading dirt-covered characters (e.g., '8' vs 'B') [INFERENCE]. | System outputs character confidence score; law enforcement manual review required for legal action [PS, INFERENCE]. |
ST_DWithin and Hidden Markov Model Map Matching) are vastly superior and more deterministic than ML models for snapping GPS pins to road network lines [INFERENCE].-- Fleet Vehicles Table
CREATE TABLE bus_fleet (
bus_id VARCHAR(32) PRIMARY KEY,
depot_id VARCHAR(32) NOT NULL,
license_plate VARCHAR(16) NOT NULL,
camera_count INT DEFAULT 3,
status VARCHAR(16) DEFAULT 'ACTIVE'
);
-- Real-Time Defect Logs Table
CREATE TABLE road_defects (
defect_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
bus_id VARCHAR(32) REFERENCES bus_fleet(bus_id),
defect_type VARCHAR(32) NOT NULL, -- Pothole, Missing_Divider, Missing_Zebra, Damaged_Sign, Waterlogging
confidence FLOAT NOT NULL,
location GEOMETRY(Point, 4326) NOT NULL,
h3_index VARCHAR(16) NOT NULL,
image_keyframe_url TEXT,
reported_at TIMESTAMPTZ NOT NULL,
status VARCHAR(16) DEFAULT 'REPORTED' -- REPORTED, VERIFIED, REPAIRED
);
-- Traffic Incidents Table (Hit & Run / Rash Driving)
CREATE TABLE traffic_incidents (
incident_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
bus_id VARCHAR(32) REFERENCES bus_fleet(bus_id),
incident_type VARCHAR(32) NOT NULL, -- HIT_AND_RUN, RASH_DRIVING
offending_plate_number VARCHAR(16),
plate_confidence FLOAT,
location GEOMETRY(Point, 4326) NOT NULL,
timestamp TIMESTAMPTZ NOT NULL,
snapshot_url TEXT,
alert_status VARCHAR(16) DEFAULT 'UNACTIONED'
);
-- Spatial Indexing for Fast GIS Aggregation
CREATE INDEX idx_defects_spatial ON road_defects USING GIST(location);
CREATE INDEX idx_defects_h3 ON road_defects(h3_index);
unsent_events.db) [INFERENCE].| Domain | Security & Governance Control | Technical / Operational Mechanism |
|---|---|---|
| Personal Data & Privacy | Anonymization of non-offending citizens and surrounding vehicle drivers [INFERENCE]. | Onboard automated face-blurring and background license plate blurring for routine road scanning logs. Unblurring occurs ONLY on triggered incident clips (e.g., hit-and-run) [INFERENCE]. |
| Location & Fleet Tracking | Protection against unauthorized tracking of public transit vehicles [INFERENCE]. | End-to-end TLS 1.3 encryption for MQTT telematics. API access restricted via JWT + Role-Based Access Control (RBAC) [INFERENCE]. |
| Authentication & Authorization | Securing edge devices and central dashboard [INFERENCE]. | Mutual TLS (mTLS) with X.509 device certificates installed on edge hardware for central broker authentication. Admin dashboard enforced with OAuth2 / MFA [INFERENCE]. |
| Data Retention & Ownership | Managing storage limits and compliance with Indian DPDP Act 2023 [RESEARCHED, INFERENCE]. | Raw rolling edge video purged after 48 hours unless flagged for evidence. PWD road defect logs retained for 3 years for municipal audit [INFERENCE]. |
| Liability & False Positives | Preventing wrongful law enforcement action based on automated ANPR alerts [PS, INFERENCE]. | System outputs ANPR confidence scores [PS]. Automated alerts serve as preliminary investigative leads for traffic police human officers, who must visually verify the keyframe before issuing an official citation [INFERENCE]. |
| Regulatory Validation | Smart City ICCC compliance [RESEARCHED]. | System output designed to adhere to Ministry of Housing and Urban Affairs (MoHUA) ICCC standards and BEL's Smart City Platform guidelines [RESEARCHED]. |
To present a compelling, failure-proof live demonstration during the SIH judging panel without requiring an actual city bus on stage, the team will execute the following hardware-in-the-loop simulation:
+-----------------------------------------------------------------------------------+
| HACKATHON DEMO SETUP |
| |
| +-----------------------------------+ +-----------------------------------+ |
| | Video Streamer (Laptop / Pi) | | Edge Processor (Jetson / Laptop) | |
| | Pre-recorded 1080p Bus Route Feed | --> | Running Real-Time YOLO + ANPR + | |
| | + Synced NMEA GPS Log Playback | RTSP| ByteTrack Pipeline | |
| +-----------------------------------+ +-----------------+-----------------+ |
| | |
| MQTT over| Live Trigger |
| Local Net| Payloads (< 5KB) |
| v |
| +-----------------------------------------------------------------------------+ |
| | Central Command Dashboard (Presenter Screen) | |
| | - GIS Map updates live bus icon along route | |
| | - Instant alert card pops up when pothole / hit-and-run occurs | |
| | - Displays plate string, confidence %, timestamp, and snapshot image | |
| +-----------------------------------------------------------------------------+ |
+-----------------------------------------------------------------------------------+
DL 3C AB 1234), computes confidence (e.g., 94.2%), and publishes an MQTT alert [INFERENCE].Given the project deadline of 30 September 2026 [PS], the implementation roadmap is structured into 4 intensive phases over the available build window:
30 Aug 2026 07 Sep 2026 15 Sep 2026 23 Sep 2026 30 Sep 2026
|--------------------|--------------------|--------------------|------------------|
PHASE 1 PHASE 2 PHASE 3 PHASE 4
Data Prep & Models Edge Pipeline & ANPR Backend & GIS Web UI Testing & Demo Prep
.engine TensorRT models achieving > 25 FPS [INFERENCE].| Test Category | Target Objective / Test Scenario | Pass Criteria | Validation Method |
|---|---|---|---|
| Functional Defect Test | Detect potholes, missing zebra crossings, damaged signs from video stream [PS]. | Detection recall > 85% across all 7 defect classes [INFERENCE]. | Evaluate model against 500 annotated test video frames [INFERENCE]. |
| ANPR Incident Test | Extract license plate number and confidence score during hit-and-run trigger [PS]. | Character recognition accuracy > 90%; plate localization recall > 95% [INFERENCE]. | Test on 100 Indian vehicle plate video clips under varied lighting [INFERENCE]. |
| Performance Test | Edge inference frame rate and CPU/GPU memory utilization [INFERENCE]. | ≥ 20 FPS per stream; GPU memory utilization < 80%; power draw < 20W [INFERENCE]. | nvidia-smi and Tegrastat profiling during 4-stream processing [INFERENCE]. |
| False Positive Test | Prevent shadows, tree leaf reflections, and road patches from triggering pothole alerts [INFERENCE]. | False positive rate < 5% on clean road control footage [INFERENCE]. | Benchmark against 1-hour clean road drive video [INFERENCE]. |
| Offline Synchronization Test | Disconnect 4G network for 15 minutes while generating edge alerts [INFERENCE]. | Zero data loss; all unsent alerts stored in SQLite and synced upon reconnect [INFERENCE]. | Network disconnect simulation via iptables / SIM toggle [INFERENCE]. |
| Spatial Deduplication Test | Send 10 duplicate pothole alerts for the same physical location from 5 different bus IDs [INFERENCE]. | PostGIS clusters 10 reports into 1 unique defect record with hit-count = 10 [INFERENCE]. | Execute automated API test script pushing cluster coordinates [INFERENCE]. |
| Risk Event | Likelihood | Impact | Severity | Mitigation Strategy |
|---|---|---|---|---|
| High False Positive Pothole Rate from Road Shadows | High | Medium | High | Multi-frame temporal persistence check (defect must appear across ≥ 3 consecutive frames) + multi-bus spatial clustering threshold [INFERENCE]. |
| Poor ANPR Accuracy on Non-Standard / Dirty Plates | High | High | Critical | Output explicit confidence score [PS]; trigger dual OCR fallback pipelines; flag low-confidence reads (< 70%) for human officer verification [INFERENCE]. |
| Edge Hardware Overheating inside Non-AC Bus | Medium | High | High | Select industrial wide-temperature fanless enclosure (-20°C to 70°C operating limit) and enforce hardware thermal throttling rules [INFERENCE]. |
| 4G Cellular Dead Zones Preventing Instant Alerts | High | Low | Medium | Onboard SQLite store-and-forward queue ensuring zero data loss upon network restoration [INFERENCE]. |
| Camera Lens Dirt / Mud Accumulation on Front Bus Bumper | Medium | Medium | Medium | Software image clarity metric (Laplacian variance score); automatically flag "Camera Obscured" alert to transit depot maintenance [INFERENCE]. |
| Component | Hackathon Option | Estimated Cost |
|---|---|---|
| Edge Hardware | Existing Laptop GPU (NVIDIA RTX 3060/4060) or NVIDIA Jetson Nano / Orin Developer Kit | Free (Student Hardware) / ~₹25,000 |
| Cameras | USB HD Webcams / Smartphone IP Camera App streams | ₹0 – ₹3,000 |
| Cloud Hosting & DB | Free-tier Cloud Instance (Render / AWS Free Tier / Supabase PostgreSQL) | ₹0 |
| GIS Tile Services | OpenStreetMap / Mapbox Free Tier | ₹0 |
| Total Hackathon Outlay | Minimal Budget | ₹0 – ₹28,000 |
| Component | Production Option | Cost Model |
|---|---|---|
| Onboard Industrial Edge Box | Industrial NVIDIA Jetson Orin Nano / Hailo-8 AI Box | ~₹45,000 per bus (One-time Capex) |
| Automotive IP Cameras (3x) | IP67 Ruggedized Wide-Angle Automotive Cameras | ~₹12,000 per bus (One-time Capex) |
| Industrial Power Supply & Enclosure | Wide-voltage DC-DC converter, wiring harness, shock mount | ~₹5,000 per bus (One-time Capex) |
| Cellular Telematics SIM | Machine-to-Machine (M2M) 4G SIM (< 500MB/month pool) | ~₹150 per bus/month (Opex) |
| Central ICCC Server Infrastructure | BEL Smart City Data Centre / Private Cloud Hosting | Shared Municipal Infrastructure [RESEARCHED] |
| Dimension | Generic Hackathon Submission | Proposed BEL Mobile Urban Intelligence Platform |
|---|---|---|
| Sensing Approach | Basic single-pothole web app reliant on manual phone camera uploads. | Passive Multi-Camera Transit Sensing: Transforms routine bus routes into an automated, city-wide mobile surveillance network [PS, INFERENCE]. |
| Edge Efficiency | Streams full video feeds to cloud or runs heavy un-optimized PyTorch models on CPU. | TensorRT INT8 Quantized Pipeline: Real-time multi-task inference (7 defect classes + vehicle counting + ANPR) on edge GPU consuming < 15W power [PS, INFERENCE]. |
| Incident Handling | Simple static map pins without actionable law enforcement data. | ANPR Incident Module: Captures hit-and-run / rash driving with plate extraction, confidence score, timestamp, GPS tag, and instant MQTT alerting [PS, INFERENCE]. |
| GIS Analytics | Basic Google Maps marker display. | Municipal Spatial Analytics: H3 spatial clustering deduplication, congestion heatmaps, route delay estimation, and automated PWD work-order generation [PS, INFERENCE]. |
| BEL Ecosystem Fit | Standalone app with no enterprise integration path. | Native ICCC Compatibility: Designed specifically to feed into Bharat Electronics Limited's (BEL) Smart City Command and Control Architecture [PS, RESEARCHED]. |
| Item Type | Description | Current Status / Assumptions Made | Validation Action Required by Student Team |
|---|---|---|---|
| [ASSUMPTION] | Onboard buses have standard 12V/24V DC auxiliary power accessible for mounting edge hardware. | Assumed standard automotive power tapping is feasible [INFERENCE]. | Validate electrical power tap specifications with local municipal transit operator (e.g., DTC / BMTC) [INFERENCE]. |
| [ASSUMPTION] | Front windshield mounting provides sufficient field of view for both road defect detection and ANPR. | Assumed 120-degree wide-angle camera covers road surface + adjacent lanes [INFERENCE]. | Test camera field-of-view angle on a test vehicle to confirm zero blind spots [INFERENCE]. |
| [NEEDS VALIDATION] | Access to live RTSP streams from existing bus CCTV cameras vs requiring dedicated IP cameras. | Unclear if existing onboard cameras expose open RTSP network streams or use proprietary closed NVRs [INFERENCE]. | Inspect camera hardware on target transit buses to verify RTSP stream accessibility [INFERENCE]. |
| [NEEDS VALIDATION] | Exact state traffic law requirements for automated ANPR evidence admissibility. | ANPR alerts serve as preliminary leads; official e-challans require manual review [INFERENCE]. | Verify local traffic police legal guidelines regarding automated camera evidence [INFERENCE]. |
The team should proceed with building Approach 3: Edge-AI Mobile Urban Intelligence Framework with Hybrid MQTT Telemetry and Central GIS Command Platform [PS, INFERENCE]. This approach fulfills every explicit requirement in SIH26124, respects bandwidth limitations, aligns with BEL's Smart City ICCC platform, and provides a highly realistic, visually impressive live demonstration for hackathon judges [PS, RESEARCHED].
Acquisition and curation of high-quality Indian road scene video datasets (IDD, Kaggle Indian License Plates) to train robust detection models that do not fail under Indian lighting and road conditions [RESEARCHED].
High false-positive rates for pothole detection caused by dark tree shadows, fresh asphalt patches, or wet road surfaces during driving [INFERENCE].
Executing real-time multi-task deep learning inference (Road Defect CV + ByteTrack Vehicle Counting + ANPR OCR Engine) concurrently on an onboard edge device within a strict 15-Watt power budget [INFERENCE].
92 / 100 [INFERENCE]. (High feasibility for a 4-6 student team; core vision, GIS, and MQTT components can be prototyped using open-source tools and datasets within 25 days [INFERENCE]).
High Confidence (0.91) [INFERENCE].
Smart City Solutions - Bharat Electronics Limited (BEL)
India Driving Dataset (IDD) - IIIT Hyderabad & Intel
Indian Vehicle License Plate Dataset - Kaggle / DataCluster Labs
CITYROVER - AI Pothole & Road Defect Detection
Pothole Detection and Analysis System (PoDAS) - arXiv
What we think about this platform, and how it can help people solve SIH problems.
" 30M+ Google Gemini tokens went into researching, analyzing, and understanding all 233 SIH problem statements to make them easier to explore and solve. "
" The platform makes SIH problem statements much easier to understand. Having everything from the problem details to the procedure in one place. "
" 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. "
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.