Hybrid spatial and channel attention in post-accident object detection

Analysing post-accident scenes using in-vehicle cameras is crucial for effective highway traffic control and enhancing accident response, road safety, and traffic flow. This contributes to a comprehensive understanding of the situation and achieves better decision-making and effective management. The accident scene report system is designed to focus on specific post-accident objects, such as crashed vehicles, involved individuals, emergency vehicles, and debris. This means that the post-accident object detection algorithm needs to handle a wide variety of objects, from large collapsed vehicles to tiny particles. It should operate in real-time on embedded boards, balancing detection accuracy and compactness to fit within the constraints of embedded computing modules. This approach aims to facilitate prompt reporting to traffic control centres. In this study, a hybrid spatial and channel attention and its pruning algorithm tailored for object detection in post-accident scenarios are proposed. This approach markedly enhances the detection performance in the unexpected accidents and malfunctioning scenes, significantly boosting the system's accuracy and processing speed. The method optimally balances the model compactness with seamless attention and pruning, making it highly suitable for real-time applications in traffic monitoring systems. The proposed seamless attention and pruning method is demonstrated using the proposed accident object detection dataset.