YOLOv5-M3:基于轻量级网络的车辆检测模型

YOLOv5-M3:a lightweight network-based object detection model for vehicles and pedestrians

针对车辆检测模型参数量大,多尺度车辆目标检测困难、重叠目标、小目标和遮挡目标容易漏检的问题.提出了一种轻量化的YOLOv5-MobileNetv3检测算法.首先,利用k-means算法来提高锚框维数聚类的效果和检测速度.其次,将YOLOv5的原始骨干网络CSPDarknet53替换为MobileNetv3进行特征提取.再次,在检测网络中,用CIOU损失函数代替GIOU损失函数,快速准确地定位图像目标区域,生成边界框,预测目标类别.使用交叉熵损失函数(cross entropy loss)作为分类损失函数.最后利用k-means算法来提高锚框维数聚类的效果和检测速度.结果表明,与YOLOv5网络相比,改进后的YOLOv5-M3检测准确率提高了5.0%,模型参数数量减少了46%,训练时间减少44.9%.改进后的YOLOv5-M3网络更小、训练时间更短、更准确地识别出目标,提高了自动驾驶系统中车辆和行人的目标检测的准确性,也为实现智能交通系统提供了一种选择.

In view of the large amount of parameters of the vehicle detection model,the detection of multi-scale vehicle objects is difficult,and the overlapping objects,small objects and occluded objects are easily missed.A lightweight YOLOv5-MobileNetv3 detection algorithm is proposed.First,the K-means algorithm is used to improve the effect and detection speed of anchor box dimension clustering.Second,replace the original backbone network CSPDarknet53 of YOLOv5 with MobileNetv3 for feature extraction.Thirdly,in the detection network,the CIOU loss function is used instead of the GIOU loss function to quickly and accurately locate the target area of the image,generate bounding boxes,and predict the target category.Use Cross Entropy Loss as the classification loss function.Finally,the K-means algorithm is used to improve the effect and detection speed of anchor box dimension clustering.The results show that compared with the YOLOv5 network,the improved YOLOv5-M3 improves the detection accuracy by 5.0%,reduces the number of model parameters by 46%,and reduces the training time by 44.9%.The improved YOLOv5-M3 network is smaller,has shorter training time,and recognizes objects more accurately,which improves the accuracy of object detection for vehicles and pedestrians in autonomous driving systems,and also provides an option for implementing intelligent transportation systems.