|
The task of environmental perception is an extremely important part of the automatic driving project, and mainstream schemes are divided into visual perception and radar perception. Visual algorithms have advantages in cost and object detection or image segmentation methods are usually used for perception. We adopt computer vision scheme to collect data and establish Transportation of Wuhan China (TOWC) datasets based on Chinese urban situation. After optimizing the selection of anchor boxes, the Yolov4 model is retrained as the object detector, which is combined with Simple Online and Real-time Tracking (SORT) algorithm and Perspective-n-Point (PnP) monocular algorithm to realize the integration of object detection, tracking. The detection and tracking methods have been evaluated against the Microsoft Common Objects in Context (MSCOCO) datasets and Multiple Object Tracking (MOT) challenge respectively. And the detection performance also has been evaluated again in our TOWC datasets. The evaluation has been conducted in challenging conditions, including occlusion, partial visibility, and under lighting variations with the mean average precision of 97.17% and the real-time speed of 30.2 FPS. The results show that our fusion detection algorithm is not inferior to other most advanced methods on the premise of excellent real-time performance. The developed model is a generic and accurate solution for traffic object detection and tracking, which can be applied to many other fields, such as automatic driving vehicle, traffic sign recognition, anomaly detection, motion analysis, or any other research areas where the traffic environmental perception is in the center of attention.
|
