Research on Small Target Detection Method of Traffic Signs Improved Based on YOLOv3

Abstract

Aiming at the problem of low accuracy and high missed detection rate of traffic signs in small target detection, this paper proposed an improved multi-branch parallel fusion feature algorithm based on YOLOv3. In order to alleviate the shortcomings of feature reduction for small targets after multiple convolutions. MPB- YOLOv3 collected and retained features in the initial stage of network transmission, pulled out three parallel branches from them, and respectively transmitted them to the last three feature extraction modules of the original YOLOv3 for dimensional splicing processing, thereby enhancing the dimensional information of detecting small and medium targets. In addition, in order to enhance the detection accuracy of small targets, this paper also designed a TD (Tiny-Detection) module, which is responsible for the feature collection and extraction of small targets in the parallel branch structure. In the experimental results, mAP0.5 was 5.35% higher than the traditional YOLOv3 and 5.00% higher than the SPP-improved YOLO algorithm. The small target detection index <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{mAP}_{\text{small}}$</tex> was 5.38% higher than the traditional YOLOv3 and higher than the SPP-improved YOLO algorithm Out of 2.03%. After a series of optimizations, the index ofmAP0.5 could reach up to 0.9485, and the index of mAPsmall could reach up to 0.3822.