Collaboratively Object Tracking With Multiple Cameras Based On Deep Learning Technologies

Since public safety has received more and more attention,video surveillance systems are becoming more and more popular in people's lives.At present,the general safety surveillance system adopts a scheme of deploying a large number of camera devices in order to expanding the field of detection range.Generally,these cameras are scattered,and the captured surveillance information lacks a fast and perfect integration mechanism,which causes the processing of monitoring data only based on video frames,and fails to effectively combine the spatial location information of the video surveillance system.In practical applications,without experienced staffs who are familiar with the monitoring area,it would be difficult for ordinary people to directly obtain the location information of an object to achieve efficient monitoring.Many major accidents may occur because of judging delaying.To overcome the shortcomings of the current video surveillance systems,this thesis focuses on three aspects,including object detection,object location and collaboratively object tracking with multiple cameras based on photogrammetry,GIS and deep learning technologies.An improved scheme is proposed to achieve object detection and tracking in multi-camera video surveillance scenarios.The main research contents and conclusions are as follows:(1)Build a spatial data model for the video surveillance system.Summarize the spatial data types of the video surveillance system and construct the spatial data model to realize the spatialization of the surveillant area.This procedure is the basis for the integration of GIS and video surveillance.(2)Explore the mapping relationship between the surveillance video images and the planar geospatial.Based on the theoretical methods of photogrammetry and computer vision,the camera models of two fields are merged to obtain a mapping model based on fixed geospatial elevation(5_?.The internal and external parameters of cameras are calculated by measuring and camera calibration to construct the mapping models for all cameras of the surveillance system,in order to realize the correlation between image information and geospatial information(3)Detect objects based on deep learning methods.In view of the particularity of the sample data source as videos,this research innovatively developed a video-based annotation tool to construct a standard dataset for model training.Making use of the trained Faster R-CNN object detection model,the detection of specific objects is achieved,and the pixel coordinate information that characterizing the object position is obtained.Combine the image and the planar geospatial mapping model of step(2)to calculate the position of the objects in the real monitoring scene.(4)Collaboratively object tracking with multiple cameras.A new method of multi-cameras cooperative object tracking based on position and visual features is proposed.Screen a camera with the best shooting angle and spatial position to undertake the object tracking task,and object handover between multiple cameras is realized.Finally,this thesis verified whether the coordinate mapping model and the object detection model are accurate and reliable,and then applied the improved method of multi-camera collaborative object tracking to the two experiments:‘generation of trajectory maps'and‘handling abnormal events'.The experimental results show that,(1)RMSE values of the sample points without distortion correction are between 4.75and 7.04,and MAE values are between 4.22 cm and 6.44 cm.RMSE values of the sample points with distortion correction are between 2.11 and 3.35,and MAE values are kept within 3 cm.Experiments show that image distortion has a serious impact on the accuracy of the result on object positioning.(2)The average accuracy mean(mAP)of the Faster R-CNN object detection method reach 90%,and the speed of object detecting a single-frame picture reaches 0.08 seconds,which achieves nearly real-time effects.(3)In the two experiments of‘generating trajectory maps'and‘handling abnormal events',the results show that abnormal phenomena such as abrupt location did not occurred when objects are across the multiple cameras'FOV,and the objects'movements that displayed on the map were in line with its real situation.In summary,the experimental results show that the improved method is reliable and efficient in the multi-camera video surveillance application scenarios.