Research On Shared Bicycle Detection And Counting Method Based On Deep Learning

As shared bicycles become more and more popular in cities,people’s"last mile"travel becomes more and more convenient.However,with the gradual increase in the number of shared bicycles,the problem of management difficulties has become more and more serious,which has brought a greater burden to urban management.With the continuous development of computer vision technology,automatic methods based on image and video processing are widely used in various fields,bringing great convenience to people’s production and life.Compared with manual management and management using positioning modules,using computer vision-based methods to manage shared bicycles has the advantages of easy access to information,large coverage and low cost.Therefore,this study introduces the automation method based on computer vision into the field of shared bicycle management,and realizes the detection and counting of shared bicycles by processing images and videos taken on urban streets.This can solve the pain points of urban management personnel when managing shared bicycles.The specific research content is as follows:（1）An object detection method for shared bicycle detection is proposed.The problems existing in the application of existing object detection methods to shared bicycle detection are analyzed.Aiming at the problem that the camera is far away from the bicycle and the shared bicycle object in image is small,which affects the detection accuracy,a feature fusion method is proposed.This feature fusion method combines the feature fusion method of top-down path,bottom-up path and skip connection,makes full use of the multi-scale information in the image,and improves the accuracy of small target detection.Aiming at the problem of large deformation of shared bicycles caused by different shooting angles,the standard convolution in Res Net is replaced by deformable convolution,and an offset is introduced to improve the feature extraction ability of the network for objects with large deformation.The effectiveness of this method was studied on the self-built shared bicycle detection dataset（SBD）and the public dataset Microsoft COCO.On SBD,this method obtained 92.7%m AP,which is 13.4%higher than before improvement,which shows that the method in this study can better complete the shared bicycle detection task.On the COCO dataset,this method obtained 42.0%AP,which is 5.8%higher than that before improvement,which shows that this method has strong generalization performance and can be applied to other goals except shared bicycles.This method can automatically detect shared bicycles from the collected images of illegal parking areas,and solve the problem that the current urban management department consumes a lot of manpower when supervising the illegal parking of shared bicycles.It can also be used as a benchmark for shared bicycle tracking networks and serve shared bicycle counting task.（2）A shared bicycle counting method is proposed.Introduce the multiple object tracking algorithm into the field of urban management to realize multiple object tracking and counting of shared bicycles.Aiming at the ID Switch problem existing in multiple object tracking for shared bicycles,a multiple object tracking method incorporating appearance features is proposed.At the same time,in order to solve the double counting problem caused by the existing counting schemes using the tracked maximum ID as the counting result,a counting algorithm based on cross-line statistics is proposed.A memento algorithm is proposed to solve double counting caused by shared bicycles repeatedly crossing the line.Aiming at the problem that repeated detection of shared bicycles affects the counting accuracy,a method based on Io U deduplication is proposed to avoid the repeated detection targets from being included in the statistics.On the self-constructed shared bicycle multiple object tracking dataset,after incorporating appearance features,MOTA increased by 3.6%,IDF1 increased by 10.4%,andID_sw decreased by 507.Twenty-three test videos were collected for quantitative evaluation of counting accuracy.The R2 of the maximum ID countint method is 0.92,and the RMSE is 7.763.The R2 of the cross-line counting method using the memo is 0.94,and the RMSE is 1.967,which is better than the maximum ID count method.The R2 of the counting method based on Io U deduplication is 0.92,and the RMSE is 1.581,which is better than the cross-line counting method using the memo.The accuracy of the automatic counting method in this study is demonstrated through experiments,and it is proved that this method can be applied to the shared bicycle counting task.