Real-Time Traffic Accident Detection Using I3d-Convlstm2d and Optical Flow

The dynamic and unpredictable nature of urban road traffic makes efficient accident detection critical for improving safety and optimizing transportation management in smart cities. This study explores advanced accident detection methods, analyzing existing techniques and categorizing various accident types, including rear-end, T-bone, and frontal collisions. We propose a novel approach using the I3D-CONVLSTM2D model, a lightweight architecture designed for smart city surveillance systems. By combining RGB video frames with optical flow data, our model effectively identifies accidents in real-time. Experimental results demonstrate the model's superior performance, achieving an 87% Mean Average Precision (MAP) compared to other approaches. Additionally, we address challenges related to dataset limitations, traffic variability, and data imbalances that impact detection accuracy. Our findings highlight the potential of integrating deep learning-based accident detection system into edge IOT devices for enhanced traffic monitoring and urban safety