Research on Small Target Detection in Driving Scenarios Based on Improved Yolo Network

Abstract

The obtainment of road condition information during driving is extremely important for a driver. However, drivers usually cannot notice multiple information at the same time, which definitely increases certain safety risks. Considering this problem, this paper designs a road information collection plus alarm system based on artificial intelligence to monitor road information. The underlying core algorithm of this system adopts the YOLO v3 network with the best comprehensive detection performance in the end-to-end network. We use this network's advantage of fast detection speed to optimize on its original basis, and propose to 'copy' part of the backbone network to build an auxiliary network, which enhances its feature extraction capability. Further, we apply the attention mechanism to the feature information fusion of the auxiliary network and the backbone network, suppress the invalid information channel, and improve the network processing efficiency. Besides, the training part of the network is optimized, and the mAP (mean Average Precision) is improved by setting the scale that meets the target to be detected. Through the test, the average test accuracy of the optimized network model reaches 84.76%, and the real-time detection speed on the 2080Ti reaches 41FPS. Compared with the previous network, the detection accuracy increases by 5.43% after optimization.