【摘要】 针对在实际交通场景中对车辆目标检测算法有占用资源小、保证实时性和准确率高的要求，提出了一种基于改进YOLOv5s的车辆目标检测算法。首先，引入GhostNet改进YOLOv5s的Backbone，降低了网络的计算量，提高了检测速度；其次，融合CBAM注意力机制，改善在各种天气、光照情况下难以被准确检测的问题；然后，使用Soft-NMS代替NMS，减少了交通拥堵等情况造成的漏检问题；最后，对改进后的算法进行了对比消融实验，验证其性能，再部署到嵌入式设备端测试。根据实验结果，改进算法在保证较高的平均精度的情况下，模型资源占用降低了34.76%，在嵌入式平台上的帧率可以达到29 frame/s，可以达到实际应用的要求。更多还原

【Abstract】 Aiming at the requirements of the vehicle target detection algorithm in the actual traffic scene,such as occupying small resources,ensuring real-time performance and high accuracy,a vehicle target detection algorithm based on the improved YOLOv5s is proposed. Firstly,GhostNet is introduced to improve the Backbone of YOLOv5s,which reduces the computation of the network and improves the detection speed. Secondly,the CBAM attention mechanism is integrated to improve the difficulty of accurate detection under various weather and light conditions. Then,Soft-NMS is used instead of NMS to reduce the problem of missing detection caused by traffic congestion. Finally,a comparative ablation experiment is conducted to verify the performance of the improved algorithm,and then it is deployed to the embedded device for testing. According to the experimental results,the resource occupancy of the model is reduced by 34.76% under the condition that the improved algorithm guarantees high average accuracy,and the frame rate on the embedded platform can reach 29 frame/s,which can meet the requirements of practical applications.更多还原