Non-motor Vehicle Detection Method And System Based On Road Monitoring Data

The report of the 19th National Congress of the Communist Party of China points out that by the middle of this century,China will be built into a prosperous,strong,democratic,civilized,harmonious and beautiful socialist modern power.Beautiful China has become the new vision of the Chinese nation,and the traffic environment is undoubtedly the vanguard of "beautiful China".By 2020,China has basically formed a new transportation system of green transportation and low-carbon transportation.With the rise of take out industry and bike sharing in recent years,the use frequency of non-motor vehicles has increased significantly.The parking problem is particularly important,and the disorderly parking of non-motor vehicles is difficult to supervise,which brings great difficulties to urban traffic management.Non-motor vehicles are easy to park because of their small size,convenient use and strong mobility.The parking is often scattered and messy.At present,the regulatory authorities need a lot of manpower and time to analyze and identify the location of non-motor vehicles in the video by accessing video surveillance,which will also lead to the lack of timeliness of supervision due to the lag of information.Therefore,although the traffic management department has invested a lot of human resources,the regulatory results still have little effect.In view of this,in order to solve the problem of traffic supervision of non-motor vehicles,this thesis establishes a non-motor vehicle detection algorithm model for monitoring data through road monitoring data and computer image recognition technology,and designs a system that can be equipped with the model,so as to realize the accurate detection of non-motor vehicles in the monitoring image,which is helpful to find the illegal parking vehicles in time and achieve early detection and early evacuation Loose.Non-motor vehicle parking detection system is of great help to alleviate traffic congestion,non-motor vehicle illegal road occupation and other situations in a timely manner.It is also conducive to strengthening the daily supervision of traffic management departments and reducing the human investment of traffic supervision departments.It is of great practical significance to improve the traffic environment and build a beautiful China.The data in this thesis come from the real road monitoring,there are great differences in the angle and height.Because there are totally different views of non-motor vehicles in the image,and the phenomenon of clustering parking is also very common.This brings great challenges to the detection task.In addition,the purpose of this thesis is to identify the non-motor vehicles in the road monitoring data,so as to assist the regulatory authorities to effectively supervise the illegal parking phenomenon in the road.For the real road conditions,there are more non-motor vehicles driven by pedestrians.Therefore,it is difficult to accurately identify the non-motor vehicles that are already in the parking state.In addition,for deep learning,it needs a lot of sample training to achieve good results,so it needs a lot of human resources in data annotation.And because of the high level of the surveillance video,the non-motor vehicle target on the edge of the image is small,and the task of detecting small target is a difficult problem in the field of target detection,which also brings some difficulties for the non-motor vehicle detection task.In response to the above problems,the research content and main work of this article are as follows:(1)This thesis presents a non-motor vehicle detection algorithm based on road monitoring data.In view of the particularity of real monitoring data,this thesis first proposes data augmentation strategies with different combinations to enhance the network’s ability to identify small targets and enhance the robustness;then,for the non-motor vehicle deformation problem from different angles in the data set,it introduces variable volume and attention mechanism,so that the network can better locate the position of motor vehicles and improve the performance In addition,through the fusion of multi-scale feature information,the network is sensitive to different scale objects in the data,which improves the ability of the algorithm to detect small targets;finally,through the improved regression and classification loss function,the overall performance of the algorithm is improved,which makes its comprehensive performance better than the mainstream target detection algorithm.(2)We design and implement a non-motor vehicle detection system.The target detection model proposed in this thesis is embedded into the system.Supervisors can view the real-time monitoring video through the system,save the monitoring video,and detect the video directly.For video data,the system can also extract video frames and convert the video into pictures for storage.And according to different needs,some modules of the network can be selected.The input traffic monitoring images can be accurately identified and displayed and saved,which improves the operability of urban traffic law enforcement personnel.The system can also easily visualize the evaluation index and understand the overall performance of the model.And the system is easy to operate,beautiful interface,for law enforcement personnel is very easy to use,and skilled use.