A Hybrid YOLO-ViT Approach for Real-Time Accident Monitoring through Automated CCTV Surveillance

Automated car accident detection from CCTV footage would have the capacity to dramatically change the face of road safety, as it pertains to faster emergency response and reduction of loss of lives. Although significant advancements in the field have been observed, most of the existing methods are based on CNN-based architectures, which leave Vision Transformers (ViTs) relatively unexplored. In this work, an attempt has been made to fill the gap by presenting a hybrid framework for real-time accident detection based on YOLO v3 and ViTs. The proposed approach first uses YOLO v3 for filtering frames containing vehicles and subsequently uses ViT in the filtered frames to detect accidents. Performance evaluation has shown that this approach of the joint usage of YOLO and ViT enhances the trustworthiness of real-time accident detection and also reduces the computational burden of using resource-intensive transformer models in the entire surveillance system.