复杂场景下对违规共享单车的细粒度检测方法

Fine-grained detection method for illegal shared bicycles in complex scenes

针对检测违规共享单车,受阴影遮挡、姿态差异和目标重叠因素影响,已有算法存在误提取、漏检和定位不准确问题,提出一种对违规共享单车细粒度检测的Bicycle-YOLO算法。针对目标受阴影遮挡的情形:(1)构建具有自适应感受野的C3_DCN模块,增强模型对共享单车识别和描述能力,缓解模型对违规共享单车细粒度检测的误提取情况;(2)依据违规共享单车存在停放姿态差异,引入上下文聚合块,提升模型对多粒度目标检测精度,减少漏检;(3)根据违规共享单车出现重叠堆放现象,使用WIOUv3损失函数,解决重叠目标细粒度特征混杂问题,精准定位目标位置。在自制违规共享单车数据集上,选取其他方法进行对比实验,结果表明Bicycle-YOLO算法的精确率、召回率、map@0.5与F1分别达到了93.4%、87.3%、91.2%、90.25%,明显优于其他方法,验证了本文方法的可行性。

Detecting illegal shared bicycles has great application value in maintaining the appearance of cities.Affected by shadow occlusion,pose difference and target overlap,the existing algorithms have the problems of false extraction,missed detection and inaccurate positioning.This paper proposes a Bicycle-YOLO algorithm for fine-grained detection of illegal shared bicycles.In view of the situation that the target is occluded by shadows,a C3_DCN module with adaptive receptive field is constructed to enhance the model's ability to identify and describe shared bicycles,and alleviate the false extraction of the model's fine-grained detection of illegal shared bicycles.According to the parking posture difference of illegal shared bicycles,the context aggregation block is introduced to improve the accuracy of the model for multi-granularity target detection and reduce missed detection.According to Shared cycling overlap pile up crime phenomenon,WIOUv3 loss function is used to solve the problem of mixed fine-grained features of overlapping targets and accurately locate the target position.On homemade violations Shared cycling data set,we do experiments on other methods,the results show that the Bicycle-YOLO algorithm precision ratio and recall ratio,map@0.5 with F1 were 93.4%,87.3%,91.2%and 90.25%,obviously superior to other methods,the feasibility of the method was verified.