Smart Railway Obstruction Detection System using IoT and Computer Vision

Technical Summary: NETRA – A Smart Railway Obstruction Detection System

Problem Statement
Indian Railways faces critical safety challenges from two primary sources: wildlife intrusions (particularly in elephant corridors) and deliberate malicious obstructions (sabotage via iron rods or boulders). Existing solutions, such as the optical fiber-based "Gajraj" system, suffer from prohibitive deployment costs ($1,000/km) and high false alarm rates, limiting their coverage to only 20 of India's 101 elephant corridors. Furthermore, many vision-based systems require continuous camera operation and GPU-level hardware, making them unsuitable for remote, power-constrained edge environments. There is a lack of unified, cost-effective, and internet-independent systems capable of real-time detection for both wildlife and obstructions.

Methodology
The paper proposes NETRA, a multi-layered intrusion detection framework designed for edge deployment on Raspberry Pi platforms. The system architecture consists of five sequential stages:

1. Probabilistic Sensor Fusion: The system employs a two-stage sensor layer to minimize false positives before engaging power-intensive components.

   • Motion Trigger: An HC-SR501 PIR sensor detects initial motion.
   • Spatial Validation: Upon PIR activation, an HC-SR04 ultrasonic sensor measures distance.
   • Fusion Logic: A probabilistic model calculates an intrusion probability ($P_{intrusion}$) by weighting PIR motion evidence ($w_{PIR}=0.4$) and distance-based spatial validation ($w_{dist}=0.6$).
   • Event-Driven Activation: A camera is activated only if $P_{intrusion}$ exceeds an empirically optimized threshold ($\tau_c = 0.65$). This approach reduces unnecessary visual processing by 52%.

2. Edge-AI Classification: Once triggered, the system classifies the threat locally using Deep Learning models tailored to the hardware:

   • Raspberry Pi Zero (Low-Cost Tier): Uses MobileNet-SSD (pre-trained on PASCAL VOC). Due to the lack of an explicit "elephant" class, it relies on a size-based heuristic (bounding box area ratio $\geq 0.25$) to infer elephants, alongside direct detection for humans and contour-based analysis for obstructions.
   • Raspberry Pi 4 (High-Performance Tier): Uses YOLOv5s (quantized to INT8, running via ONNX Runtime). This model is fine-tuned on a custom dataset of 2,500+ railway images, enabling native multi-class detection (elephants, humans, obstructions) without heuristics.

3. Alert Transmission: Confirmed threats are transmitted via LoRa (868 MHz) to an ESP32 microcontroller on the approaching train. The system uses an adaptive spreading factor (SF7–SF12) to ensure reliable communication over 1–2 km without cellular infrastructure.

Key Contributions

• Unified Detection Framework: NETRA is the first proposed system to simultaneously detect wildlife (specifically elephants) and malicious obstructions within a single, internet-independent architecture.
• Probabilistic Sensor Fusion: The integration of PIR and ultrasonic sensors with a tunable threshold ($\tau_c = 0.65$) achieves 95% detection accuracy with zero false alarms, significantly outperforming binary fusion methods (85% accuracy).
• Cost-Effective Edge Deployment: The system reduces deployment costs by 75% compared to the Gajraj system ($247/km vs. $1,000/km). It offers a tiered deployment strategy:
  • Tier-1 (Critical): Pi 4 + YOLOv5 ($100/km) for high-risk elephant corridors.
  • Tier-2 (Standard): Pi Zero + MobileNet-SSD ($20/km) for moderate-risk sections.
• Energy Efficiency: By activating the camera only upon confirmed intrusion, the system reduces energy consumption by 62.8% compared to continuous monitoring, enabling solar-powered operation.

Experimental Results
The system was evaluated across 113 motion events in a semi-outdoor environment over one week:

• Detection Accuracy: The probabilistic fusion achieved a 95% detection rate with 0% false alarm rate.
• Classification Performance:
  • Pi 4 (YOLOv5): Achieved an 83.5% F1-score for elephant detection, a 5.6× improvement over the Pi Zero's heuristic approach. It also achieved 92% F1-score for humans and 100% for obstructions.
  • Pi Zero (MobileNet-SSD): Achieved only a 14.8% F1-score for elephants due to size-based ambiguity between elephants and large cattle (buffaloes), rendering it unsuitable for critical elephant corridors but viable for general human/obstacle monitoring.
• Latency and Reliability: The end-to-end latency (detection to alert) was 2.4 seconds on the Pi 4 platform. LoRa communication achieved 100% packet delivery across 1–2 km in field trials.
• Field Validation: In a 7-day deployment, the system correctly identified 42 human intrusions, 15 animal incursions, 8 elephant crossings, and 12 track obstructions, transmitting 77 dangerous alerts with zero false alarms.

Significance and Claims
The paper claims that NETRA provides a practical, scalable solution for remote railway safety where existing infrastructure is absent or too expensive. By combining sensor fusion with tiered edge computing, it addresses the dual threats of wildlife and sabotage while maintaining a low false alarm rate—a critical requirement for railway operations. The authors emphasize that the system's ability to operate without internet connectivity and its 75% cost reduction make it a viable alternative to current fiber-optic solutions, potentially enabling coverage of all 101 elephant corridors in India.

Limitations Acknowledged
The authors note several constraints:

• The Pi Zero's heuristic approach is insufficient for elephant detection in critical zones.
• The ONNX deployment on Pi 4 introduces systematic misclassification of cows/buffaloes as humans due to backend parsing differences.
• Nighttime detection is currently unsupported as the camera lacks infrared capability.
• Field validation was conducted in a semi-outdoor environment rather than active railway tracks due to safety constraints.