Vehicle trajectory recognition based on video object detection

摘要

Highway always has high-speed traffic flow, which means that frequent lane changes will easily cause large risk. So, lane changes are often not allowed on complex sections, such as tunnel, long downhill section, etc. Vehicle trajectory recognition from the video can help the administration monitor and analyze the movement of the vehicles. In this paper, we choose a one stage object detection network called Yolo to detect vehicles from the surveillance video camera. Data augmentation, focal loss, and synchronized batch normalization are applied to improve the performance of detector. After successful vehicles detection, a vehicle box matching method based on IOU is applied to identify whether a detected vehicle is a recorded vehicle or new one. The results show that the object detection and tracking method can detect and track vehicle is stable, the trajectory recognition achieves high reliability.