Lightweight Detection Network of Trunk Contents

Abstract

During the annual vehicle review, real-time identification of oil-to-gas vehicles is one of the most important contents. Aiming at the characteristics of objects in the trunk of such vehicles, a Trunk-YOLO lightweight real-time detection network is proposed. Firstly, by streamlining the number and structure of CSP modules, and cutting the convolutional layer to perform feature extraction, the cost of feature extraction will be reduced while modifying the activation function. Secondly, a faster K-means algorithm is proposed to perform target candidate frames for the features of the trunk object cluster analysis of numbers. Finally, the YOLO-Head module performs classification and regression, and outputs the prediction results. The method proposed in this paper is tested on the self-made dataset TRUNK-Dataset and compared with the classic lightweight target detection network. The test results show that the network proposed in this paper has achieved good results in detection speed and accuracy.