Vision-Based Automated Road Accident Detection and Emergency Alert System using Lightweight LSTM and Simulated Sensor Fusion

Abstract

Road traffic accidents remain a significant publicsafety problem and one of the leading causes of death worldwide. Delayed emergency response can substantially worsen accident outcomes, particularly in regions where timely medical assistance is not readily available. Conventional accident-detection systems rely on in-vehicle sensors such as accelerometers and gyroscopes, which are costly to deploy across large vehicle fleets and are prone to false alarms during hard braking or travel on rough roads. This paper proposes a fully software-based accident detection and alert system that does not depend on physical inertial sensors. Traffic surveillance video is processed with a fine-tuned YOLOv5-small detector for vehicle detection and tracking. We employ a two-layer LSTM network with 64 hidden units per layer and a Leaky ReLU activation, operating over a temporal motion-modelling window of 10 frames. Virtual acceleration and angular-velocity signals are derived from frame-to-frame pixel displacement and fused with visual anomaly and temporal scores through a calibrated weighted fusion model. Evaluated on 2,916 clips drawn from the CADP and CCD benchmark datasets, including 1,100 confirmed accident events, the proposed system achieves 93.8% accuracy, 92.4% precision, 94.1% recall, a 93.2% $F_{1}$-score, and a mean confirmation latency of 4.5 seconds. In this revised version, we further analyse multi-object tracking under occlusion, report the computational complexity and per-stage inference time of the pipeline, discuss reliability under varying lighting and weather conditions, examine robustness to camera viewpoint and scale variation, justify the YOLOv5-small design choice, and outline multi-camera and smart-city deployment considerations. The results confirm the effectiveness of the proposed hardware-free framework for intelligent transportation systems and smart-city surveillance applications.