Object Detection in Complex Road Scenarios: Improved YOLOv4-Tiny Algorithm

Abstract

This paper proposes an object detection algorithm based on YOLOV4-Tiny, which can improves the recognition of small objects in street scenario, as well as the robustness in the extreme climate, e.g., the rainy or foggy days. Specifically, as for the small objects in the street scenario, we first conduct the up-sampling in the feature map obtained by 8x down-sampling, and then fused it with the feature map obtained by 4x down-sampling. Meanwhile, as for the images in the extreme climate, we use the sharpness algorithm based on secondary blur (ReBlur) to characterize the image blurriness, where the blurred images will be restored a dark channel prior algorithm. The simulation results indicate that the proposed algorithm can improve the recognition of the objects in the complex street scenario, where the mean average precision (mAP) is increased by 4.13%.