基于YOLOv5s改进的无人机航拍图像车辆检测模型

Improved UAV Aerial Image Vehicle Detection Model Based on YOLOv5s

针对无人机航拍图像车辆检测任务中存在车辆遮挡严重、小尺度目标多、背景信息复杂、误检漏检情况严重等问题,提出一种基于YOLOv5改进的车辆目标检测模型。首先,增加一个小目标特征检测层,增强对浅层特征图中有效位置特征信息的复提取,从而缓解因深层卷积导致密集小目标特征信息的缺失问题。其次,在Neck中使用GSConv卷积和VOVGSCSP模块,对模型进行轻量化同时提高检测精度。再次,使用Mish作为全局激活函数,提高特征信息在深层网络中的传播和表达能力。然后,为了模型对检测目标的定位精度,使用EIoU作为回归框定位损失。最后,在Backbone中引入Transformer模块,增强模型感受野,提高对关键点信息的提取能力,增强模型抗干扰能力。实验结果表明,最终改进模型的平均检测精度(mAP)达到了83.8%,比原始YOLOv5s模型提高了5.5%,对小目标检测精度明显得到提升。

Aiming at the problems of serious vehicle occlusion,many small-scale targets,complex background information,and serious false detection and missed detection in UAV aerial image vehicle detection tasks,this paper proposes a vehicle target detection model based on YOLOv5.Firstly,a small target feature detection layer is added to enhance the complex extraction of effective location feature information in the shallow feature map,so as to alleviate the problem of lack of dense small target feature information caused by deep convolution.Secondly,GSConv convolution and VOVGSCSP modules are used in Neck to lighten the model and improve the detection accuracy.Thirdly,Mish is used as the global activation function to improve the propagation and expression ability of feature information in the deep network.Then,for the model’s positioning accuracy of the detection target,EIoU is used as the regression box to locate the loss.Finally,the Transformer module is introduced in Backbone to enhance the model receptive field,improve the extraction ability of key point information,and enhance the anti-interference ability of the model.Experimental results show that the average detection accuracy(mAP)of the final improved model reaches 83.8%,which is 5.5%higher than that of the original YOLOv5s model,and the detection accuracy of small targets is significantly improved.