FBS_YOLO3 vehicle detection algorithm based on viewpoint information

Abstract

The FBS_YOLO3 vehicle detection algorithm is a novel solution to the challenge of detecting vehicles in unstructured road scenarios with limited warning information. This algorithm builds upon the YOLOv3 model to deliver advanced multi-scale target detection. Firstly, FBS_YOLO3 incorporates four convolutional residual structures into the YOLOv3 backbone network to obtain deeper feature knowledge via down-sampling. Secondly, the feature fusion network is improved by implementing a PAN network structure which enhances the accuracy and robustness of viewpoint recognition through top-down and bottom-up feature fusion. Lastly, the K-means clustering fusion cross-comparison loss function is utilized to redefine the anchor frame by employing a K-means fusion cross-ratio loss function. This innovative approach solves the issue of mismatching the predetermined anchor frame size of the YOLOv3 network. Experimental results demonstrate that FBS_YOLO3 on a self-built dataset can improve mAP by 3.15% compared with the original network, while maintaining a quick detection rate of 37 fps. Moreover, FBS_YOLO3 can accurately detect vehicles, identify viewpoint information, and effectively solve the warning information insufficiency problem in unstructured road scenarios.