YOLOX-B: A Better Yolox Model for Real-Time Driver Behavior Detection

Abstract

In the coal transportation scene, the object detection model proposed for the driver behavior detection task generally has the problems of inaccurate positioning and difficult detection of small objects, we propose a new model YOLOX-B, which introduces a serialized atrous spatial pyramid pooling structure (S-ASPP), obtains different sizes of receptive field information through serialized atrous convolution, solves the problem of information loss in max-pooling, and maximizes the efficiency of atrous convolution. Meanwhile, by introducing a lightweight feature reorganization module based on transposed convolution, adaptively predicting the up-sampling kernel weight, the model can better complete pixel recovery in a weighted way and improve the detection accuracy of small objects. The experimental results on the publicly available PASCAL VOC 2012 dataset and the self-built driver behavior dataset demonstrate that the YOLOX-B maintains the same inference speed as YOLOX-S, and its mean Average Precisions(mAPs) are improved by 4.4% and 0.8%, respectively.