改进YOLOv5网络模型的车位检测方法

Development of parking slot detection method based on an improved YOLOv5 model

针对目前车位检测方法效率低的问题,提出一种轻量级车位检测方法OG-YOLOv5.首先,基于YOLOv5网络添加车位分隔线方位回归分支,实现车位方位的准确预测,可直接根据网络预测结果推断完整车位信息;其次,通过检测尺度裁剪、 Ghost模块重构网络实现模型轻量化;最后,通过在网络主干中引入ECA注意力机制、优化损失函数提高目标预测精度.通过对比实验,结果表明,所提OG-YOLOv5网络的mAP@0.5达到98.8%,模型参数量和计算量仅为原模型的32.0%和28.3%,在GPU和CPU上的检测时间分别减少16.2%和28.1%,车位检测准确率和召回率分别达到97.75%和96.87%.

Aiming to improve working efficiency of parking slot detection,a lightweight parking slot detection method as OG-YOLOv5 is proposed in this investigation.Firstly,a regression branch of parking space divider orientation is added to obtain accurate prediction value of parking position on basis of YOLOv5 network.Then,the complete parking space information is inferred directly from the prediction results.Secondly,the proposed model is lightened by cropping detection scale and reconfiguring Ghost network.Furthermore,the ECA attention mechanism is introduced in the network backbone and loss function is optimized to improve the accuracy of detection result.Experimental results reveal that the mAP@0.5 of OG-YOLOv5 is 98.8%,and the number of parameters and the floating-point operations is only 32.0%and 28.3%of the corresponding results of original model,and the detection time on GPU and CPU is reduced by 16.2%and 28.1%respectively by comparison with the original model.Moreover,the performances of parking slot detection accuracy and recall ratio of the suggested method achieve 97.75%and 96.87%,respectively.