Aiming at the problem of insufficient cross-scene detection performance of
current traffic target detection and recognition algorithms in automatic
driving, we proposed an improved cross-scene traffic target detection and
recognition algorithm based on YOLOv5s. First, the loss function CIoU of
insufficient penalty term in the YOLOv5s algorithm is adjusted to the more
effective EIoU. Then, the context enhancement module (CAM) replaces the original
SPPF module to improve feature detection and extraction. Finally, the global
attention mechanism GCB is integrated with the traditional C3 module to become a
new C3GCB module, which works cooperatively with the CAM module. The improved
YOLOv5s algorithm was verified in KITTI, BDD100K, and self-built datasets. The
results show that the average accuracy of mAP@0.5 is divided into 95.1%, 72.2%,
and 97.5%, respectively, which are 0.6%, 2.1%, and 0.6% higher than that of
YOLOv5s. Therefore, it shows that the improved algorithm has better detection
and recognition accuracy and better generalization performance in
cross-scene.
