Accident Detection with Capsule Network by Quantum Trace Distance-based Routing

The continuous increase in vehicle volume has led to a high risk of traffic accidents, making CCTV image-based automatic detection an urgent necessity for public safety. However, visual data on accidents exhibit complex spatial-temporal dynamics and are often affected by low recording quality, necessitating a more robust analysis method. We propose Quantum Trace Distance-based Routing (QTD-routing), a modification of the dynamic routing algorithm in the Capsule Network method, which replaces the dot product-based agreement mechanism with the trace distance between quantum capsule states. This approach provides a more stable and discriminative similarity measurement while maintaining spatial hierarchy and feature relationships. The implementation of QTD-routing for accident detection from CCTV images demonstrates superior performance across all evaluation metrics, utilizing hundreds of training images. Our model achieves accuracy with a total of 14,827,856 parameters without sacrificing precision or sensitivity, and surpasses existing state-of-the-art methods. Our research contributes to the development of CapsNet by integrating quantum computing principles, which have the potential to enhance the performance of future visual detection systems.