轻量化的YOLOv8锥桶检测算法研究OA北大核心CSTPCD

This paper proposes a modified YOLOv8n algorithm to address the high computational complexity and low accuracy in the detection of cone barrels in existing unmanned formula racing cars.The algorithm achieves high detection accuracy and requires fewer model parameters.To improve the algorithm,the Stem module and EfficientNet-Lite network structure are first introduced to replace the YOLOv8 backbone network.Since YOLOv8's detection heads are decoupled parameters and account for 40%of the total,a lightweight detection head structure is designed to reduce the number of parameters in the model.Finally,a high-resolution feature map P2 with a downsampling factor of 4 is added to detect small targets.Our experimental results show the improved YOLOv8 algorithm improves the average accuracy index from 90.1%to 93.8%on the dataset compared to the original YOLOv8n and reduces the parameter count from 3.0 M to 1.37 M and the computational complexity from 8.1 GFLOPs to 4.7 GFLOPs.A vehicle test reveals it effectively reduces the missed detection of cone barrels and cuts the model memory by 49%.