A Review on Semantic Understanding of Road Incidents Based on Vehicle-mounted Vision Systems

Abstract

Road incidents pose serious threats to driving safety. In recent years, thanks to rapid advances in vehicle-mounted hardware and artificial intelligence technologies, it has become possible to apply these technologies for real-time road monitoring. This paper reviews methods for detecting and semantically understanding road incidents based on vehicle-mounted vision systems, focusing specifically on the applications and challenges of object detection and semantic segmentation in traffic scenarios. The paper introduces commonly used object detection models (e.g., YOLO, Faster R-CNN) and semantic segmentation models (e.g., DeepLab, PSPNet), along with their optimization methods. These optimization methods propose various improvement strategies such as attention mechanisms, feature fusion, and lightweight design to balance detection accuracy, real-time performance, and computational resources in complex traffic environments. Finally, the paper discusses future research directions, including improving model robustness under extreme conditions, multimodal data fusion, and hardware acceleration, aiming to promote broader applications of vehicle-mounted vision systems in intelligent transportation.