面向车辆目标检测的毫米波雷达和相机融合方法OA

To address the challenges of poor recognition performance with single sensors in vehicle detection and the association errors caused by vehicle occlusion between different sensors,a vehicle detection method based on the fusion of onboard millimeter-wave radar and vision is proposed.First,an improved YOLOv8n_M model is utilized to process image data.This model enhances target features by incorporating a SimAM attention mechanism into the Neck and Head sections of the original YOLOv8n model.Additionally,Wise-IoU v1,featuring a dynamic non-monotonic focusing mechanism,is employed as the loss function to improve boundary box regression performance,while a small target detection layer(P2)is added to better detect small vehicles.Concurrently,radar data is analyzed and preprocessed to filter out valid radar targets,which are then tracked using a Kalman filter algorithm.The camera and radar data are aligned temporally and spatially.Finally,the overlap rate of detection boxes and the normalized Euclidean distance of their center points are calculated to construct an association matrix,which,combined with the Hungarian algorithm,facilitates data matching and outputs fused targets.Experimental results demonstrate that,in both the BDD100K and self-made datasets,YOLOv8n_M outperforms the original YOLOv8n with an increase in mAP50 by 4.7%and 3.6%respectively,and an improvement in mAP50～95 by 2.9%and 5.4%respectively.In complex traffic scenarios,the association accuracy of the proposed association algorithm is improved by 4.66%and 2.91%compared to the traditional nearest neighbor and global nearest neighbor algorithms,respectively.The detection rate of the fusion detection reaches 88.09%,which is higher than that of a single sensor,enabling real-time and accurate detection of vehicle targets.