【摘要】 针对分类化检测处置汽车违规停放的交通管理需求，该文提出一种复杂场景下违规停放汽车的细粒度检测方法。该算法在YOLOv7的基础上，根据目标间相互重叠的现象，使用MPDIOU损失函数来优化边界框的几何一致性，提高目标的精确定位能力。针对目标尺度差异大的问题，在颈部网络引入MSDA模块，提升了模型对于不同尺度目标的特征感知力。面对目标区域遮挡问题，将RFCAConv卷积融入ELAN模块，以增强模型对目标关键特征的捕获能力。最后，在违规停放汽车数据集上进行对比实验，结果表明MRM＿YOLO算法的精确率、召回率、mAP@0.5与F₁分别达到了91.2%、86.2%、85.4%、88.6%,明显优于其他方法，验证了本方法的可行性。更多还原

【Abstract】 This article proposes a fine-grained detection method for illegal parking of vehicles in complex scenarios in response to the traffic management needs of classified detection and disposal. This algorithm is based on YOLOv7 and uses the MPDIOU loss function to optimize the geometric consistency of the bounding box and improve the precise positioning ability of the target, taking into account the phenomenon of overlapping between targets. In response to the problem of significant differences in target scales, the introduction of MSDA modules in the neck network has improved the model’s feature perception for targets of different scales. Faced with the problem of occlusion in the target area, RFCAConv convolution is integrated into the ELAN module to enhance the model’s ability to capture key features of the target. Finally, comparative experiments were conducted on the dataset of illegally parked cars, and the results showed that the Precision, Recall, and mAP@0.5 Compared with F₁,it achieved 91.2%,86.2%,85.4%,and 88.6%,which is significantly better than other methods, verifying the feasibility of this method.更多还原