基于目标检测和语义分割共享单车类别与违停检测  

Sharing bicycle’s categories and parking violation detection based on object detection and semantic segmentation

作　　者：严广宇 刘正熙[1] 熊运余[1] 李征[1] 赵逸如 Yan Guangyu;Liu Zhengxi;Xiong Yunyu;Li Zheng;Zhao Yiru(School of Computer,Sichuan University,Chengdu 610000,China)

机构地区：[1]四川大学计算机学院,成都610000

出　　处：《计算机应用研究》2020年第10期3175-3179,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61471250)。

摘　　要：目前道路违规事件检测多在固定摄像头下人工框定区域进行检测,但人工框定工作量大,并且摄像头转动会使得框定区域失效。针对此问题,率先提出一种目标检测与语义分割相结合的违停检测方法。该方法首先使用目标检测Faster R-CNN,采取迁移学习、多阶段训练等方法建模,提取共享单车的类别与检测框位置信息。再使用group normalization改进语义分割DeepLab v3+网络模型,提高其在小batch size下训练的模型精度,用于分割图像获得道路的语义和区域信息。最后综合两部分信息,根据单车检测框内不同道路区域所占比例判定共享单车是否属于违规停放。实验结果表明,该方法对共享单车类别的mAP为72.36%,对共享单车违规停放的平均检测率为89.11%,适用于真实城市道路监控环境中。At present,the detection of road violation event is based on the artificially selected area under the fixed camera for object detection.However,the artificially selection area has a heavy workload,and as the camera rotates,these area will be invalid.Aiming at this problem,this paper first proposed a method of violation detection combining object detection and semantic segmentation.Firstly,this paper used the transfer learning,multi-stage training schedule to train the Faster R-CNN model to extract the categories of the shared bicycles and the bounding boxes position information.Then this paper used group normalization to modify the semantic segmentation DeepLab v3+network model,improved the model precision of training under small batch size,and obtained the semantic and regional information of the road scene.Finally combining two parts of the information to determine whether the sharing bicycle was illegally parked according to the proportion of different road areas in the bicycle detection bounding boxes.The experimental results show that the mAP of the sharing bicycle detection is 72.36%,and detection accuracy of the sharing bicycles illegal parking is 89.11%,which can be applied to the actual road monitoring environment.

关 键 词：共享单车 目标检测 语义分割 违停检测 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]