军事知识服务平台
Research on autonomous driving object detection based on CM-YOLOv8
摘要 To tackle the dual challenges of low accuracy in small object detection and high computational complexity in current autonomous driving algorithms,this paper proposes an enhanced object detec-tion framework based on the you only look once version 8(YOLOv8).First,the original C2f module is substituted with a lightweight GhostBottleneck to minimize computational overhead.Second,a multi-scale fusion mechanism is integrated into the neck network to optimize the aggregation of deep and shallow features.Finally,addressing the insufficient utilization of spatial information in standard fusion,we introduce spatial position encoding to perform weighted feature fusion.This strategy sim-ultaneously compresses features to reduce computational load and boosts prediction accuracy.Exper-imental results on the Karlsruhe Institute of Technology and Toyota Technological Institute(KITTI)dataset demonstrate that the proposed model achieves a precision of 92.4%,mAP@50 of 91.9%,and mAP@50∶95 of 73.9%.Compared to the baseline YOLOv8,these metrics represent improve-ments of 3.2%,7.4%,and 15.7%,respectively.Furthermore,the model outperforms main-stream detectors such as YOLOv5,YOLOv11,and real-time detection Transformer(RT-DETR-L),successfully balancing a lightweight design with superior detection performance for autonomous driv-ing scenarios.
年,卷(期): 2026, 32(3)
机标分类号: F279.23 TP393.08 R917
在线出版日期: 2026年9月28日