Object Recognition Algorithm for Complex Scenes Based on Improved YOLO v3

Abstract

YOLO v3 is widely used in industry because of its high detection accuracy and speed, but there is a problem that it can only output accurate position coordinates and cannot predict the localization uncertainty of bbox. To solve this problem, an improved YOLO v3 algorithm is proposed. By increasing the output of position parameters and predicting localization uncertainty of bbox with Gaussian modeling to remove the boxes with high bbox uncertainty in the detection process. A new Localization loss function is designed on the basis of increasing the output of bbox coordinates. Batch Normalization layer and Convolution layer are combined to reduce the use of video memory space and improve network performance. The experimental results show that the mAP50 of the improved YOLO v3 algorithm in the helmet wearing test set is improved by 7.99%.