关键词: YOLOv8n, 车辆检测, 全局注意力机制, 损失函数, 主干网络

Abstract: With the development of autonomous driving technology,object detection has become a prerequisite for safe driving in autonomous driving due to its precise perception of the environment,real-time processing,and robustness.In crowded traffic scenarios where vehicles frequently obstruct one another,occlusion is still a significant challenge. Existing improved algorithms had exhibited certain deficiencies in occlusion perception.To address these issues. Firstly,the study proposed a new backbone network module,C2fCIBG,to deal with poor information extraction. Then,the GAM attention mechanism and CIB deep convolution were added to the new module to better capture key features and improve extraction. Finally,since CIOU was not robust enough,the loss function was changed to MPDIOU. Changing the loss function helped the model adjust the prediction box′s position and size more accurately, improving feature extraction and classification in the target area. Thus,it met the needs of vehicle occlusion detection in autonomous driving,the KITTI dataset was chosen for the experiment. The improved algorithm was compared with the traditional YOLOv8n algorithm.Evaluation was based on three key metrics:detection accuracy,recall rate,and mean average precision.The results showed increases of 0.87 %,2.0 %,and 2.7 % respectively,which proved that the improved algorithm was viable for vehicle occlusion perception.

Key words: YOLOv8n, vehicle detection, Global Attention Mechanism (GAM), loss function, backbone network