Road Traffic Accident Detection Using Autoencoder-Based Models: A Performance Analysis of CAE and VAE

Road traffic accidents are a major public health and safety issue worldwide, and thus there is a need to design systems that can automatically detect accidents in time to provide emergency response and manage traffic. In this paper, a detailed performance analysis of Convolutional Autoencoder (CAE) and Variational Autoencoder (VAE)-based architectures for the detection of road traffic accidents in surveillance videos is presented. Unlike supervised approaches that require extensive labelled datasets, autoencoders learn normal traffic patterns and identify accidents as anomalies through reconstruction error analysis. We evaluate both models on two benchmark datasets: the IITH road accident dataset from Hyderabad, India, and the UCF-Crime dataset's road accident subset. Our experimental results demonstrate that CAE consistently outperforms VAE across multiple metrics, achieving 86.78% accuracy on IITH dataset compared to VAE's 83.42% and achieving 81.63% accuracy on UCF-Crime Dataset compared to VAE's 78.56%. The deterministic latent representations of CAE prove more effective for accident detection than VAE's probabilistic modelling, though VAE's uncertainty quantification offers potential advantages for confidence estimation. Furthermore, we provide an analytical discussion on the implications of reconstruction-based anomaly scoring for downstream applications such as traffic video summarization, where high reconstruction errors can serve as saliency indicators for keyframe selection. This study contributes empirical evidence for selecting appropriate autoencoder architectures in intelligent transportation systems and highlights the trade-offs between deterministic and probabilistic latent space modelling for real-world traffic surveillance applications.