Research On Object Detection Method In Driving Scenario Based On Improved YOLOv4

Abstract

The detection of vehicles and pedestrians is driving ahead has been a hot topic in the field of computer vision research. Although the current field of object detection developed rapidly, there are still problems such as the limited effect and slow detection speed of small object detection. Based on YOLO v4, this study proposed an algorithm that takes into account both detection speed and detection accuracy. First, MobileNet v1 was applied to improve the slow detection speed and reconstruct the original feature extraction network. Secondly, the deleted original 13 x13 prediction head and the added <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$104\times 104$</tex> prediction head were applied to resolve the general detection effect of small targets. Finally, the K-means algorithm was applied to cluster analysis data set and generate the initial anchor box of the network. The improved algorithm results showed that the average accuracy (mAP) reached 90.32%, and detection speed can reach 35FPS. The mAP was reduced by 2.66% compared with the YOLO v4 algorithm, but the model size was only 23.70% of the original one, and the detection speed was 1.66 times that of YOLO v4. Therefore, the improved algorithm is a reliable algorithm that can be applied to target detection in driving scenarios.