Research On Object Tracking Based On Multiple Collaborative Camears

With the rapid development of the information age,the need of understanding the semantic information of images is increasing.As object tracking is basis for understanding deep semantic information,it has also become an attractive and challenging research direction in the field of computer vision.Object tracking technology has wide application space in intelligent transportation,security monitoring,human-machine interaction and other fields,and has high research significance and practical value.In the real world,object tracking based on the single camera has such problems as limited tracking field of vision,single tracking angle,hardly solving target occlusion.In contrast,the object tracking system based on multi camera can solve these problems well and is becoming a hot spot of research.The research content of this thesis mainly includes three aspects,which are object detection,single camera target tracking and multi camera coordination object tracking.Object detection is the precondition and foundation of the follow-ups.The single camera object tracking is based on the object detection to track the target in the range of camera.The object tracking of multi camera is based on the object tracking of a single camera and fuses information of multiple cameras to achieve a long,multi angle tracking of the object.The main contents of this thesis are as follows:(1)The popular object detection algorithm Faster RCNN,which is based on the convolution neural network,is used to detect the object.The Faster RCNN algorithm is compared with the DPM(Deformable Parts Model)algorithm,which is based on the artificial selection feature,and the experiment is carried out to compare two algorithms.The experimental results show that compared with the artificial selection feature,the feature obtained by the convolutional neural network method is more accurate,and the Faster RCNN algorithm can obtain better detection results under the condition that the background is complex or the target is occluded.(2)On the basis of object detection,an online object tracking framework is proposed.The framework uses Kalman filter to predict and update the status of the object.For single object tracking,the similarity vector is constructed by calculating the similarity between the detected objects and the tracking object,so that the detection object and the tracking object can be associated to achieve the object tracking.For multi-object tracking,we need to calculate the similarity between the detected objects and the tracking objects,and construct the incidence matrix,and establish the one-toone correspondence between the detected object and the object being tracked by Hungary algorithm.In order to establish an accurate and robust model for object effectively,this thesis proposes an object model that integrates the overlap rate between objects and the improved spatiograms of adaptive scale.In multi-object tracking,this thesis uses the relative position between objects to put forward the occlusion prediction method to solve the mutual occlusion problem between objects.(3)In view of the object matching in the multi camera environment,two different matching algorithms are studied for different situations.When the camera has the overlap with other cameras,the SIFT feature are used to calculate the homography matrix,and then fuse the information of two cameras,at the same time,we can solve the problem of object occlusion.In view of the non-overlapping field between two cameras,the method of combining SIFT feature and the improved spatiograms feature are combined to model the object to handle the matching.The object tracking in the multi-camera environment proposed in this thesis can achieve a wide range of target tracking and can better solve the object occlusion problem.