Research On Key Issues Of Multi-Object Relay Tracking In Non-overlapping Cameras System

With the promotion of national strategies for smart cities and safe cities, more and more cameras are installed in some important places in our cities, such as schools, shopping malls, railway stations, hospitals, communities, and so on. However, at present, each installed camera is just playing a role in monitoring individually, and thus how to let these separated cameras combine together to form a monitoring network and track objects cooperatively in a multi-camera environment has been a hot topic in the field of intelligent visual surveillance. However, there often exists some blind spots in the current visual surveillance systems with non-overlapping views, which makes it difficult to achieve the cooperative object tracking among multiple cameras, and thus how to eliminate the visual differences between different cameras and overcome the impact of the spatial and temporal uncertainty of blind spots on object tracking are two big concerns in this dissertation. For this purpose, this dissertation discusses object detection and tracking in a single camera, object recognition between different cameras and object association based on the topology of a camera network in detail. The main research contents and innovative contributions of this dissertation are as follows:(1) Considering a single camera and a fixed monitoring scene, this dissertation presents a detection algorithm for multiple moving objects based on VIBE and HOG feature classifications and a multi-object tracking algorithm based on two-phase distance matching. In order to reduce the effect of illumination change and partial occlusion, this dissertation first obtains the possible area of each object based on the improved VIBE algorithm. In addition, the SVM classifier with HOG feature is used to segment moving objects accurately for achieving object detection. Moreover, the bi-direction optimal matching is executed between foreground areas and objects according to the idea of forward-backward error, and objects that are not matched successfully will be eliminated or increased for achieving object tracking. Finally, the proposed algorithms are tested on a large number of standard and autodyne video sets and the experimental results show that the object detection algorithm has a relatively high accuracy and the object tracking algorithm has a good robustness.(2) An object re-identification algorithm for single sample based on multiple features fusion and an object re-identification algorithm for large samples based on the Jensen-Shannon (JS) kernel are proposed. For single sample, this dissertation first extracts the color histogram features and texture features of pedestrian images and builds the corresponding binary classifiers. In addition, different features with different kernel function are assigned for training, and the optimal weights of different kernel function can be obtained by the multiple kernels learning algorithm. Moreover, typical negative samples are selected by the Bhattacharyya coefficient based affinity propagation clustering algorithm to solve the sample imbalance problem in which negative samples are far more than positive samples in the process of training. As for multiple samples, this dissertation firstly verifies that the JS kernel has a good adaptability to different color histogram features and uses it to map each image into a high-dimensional space for making the same class close and different classes apart, Besides, for finding the direction of projection, the scatter of the samples can be discriminated and analyzed by the LFDA algorithm. Finally, the feasibility of the proposed algorithms is demonstrated by a larger number of tests based on the three public benchmark datasets, such as VIPeR, i_LIDS and ETHZ, as well as the autodyne datasets.(3) In order to reduce the mismatches, an object association algorithm based on Bayesian model and discrete particle swarm optimization (DPSO) is developed. This dissertation first estimates the topological relation among online cameras in a camera network. Besides, the Bayes model is adopted to fuse the object re-identification results and the topological relation of online cameras and an evaluation function is constructed to achieve the optimal path selection. Moreover, DPSO is introduced to search for the optimal path set of all objects for decreasing the computational complexity. Finally, the simulation experiments in a large-scale visual surveillance system show that the proposed algorithm has a good timeliness and can give the most likely trajectory of each object in a certain period of time.