Object Detection, Tracking And Re-Identification In Multi-Camera

Video surveillance systems is widely used and developed in many parts of human life that’s because the huge step of the technology in the last two decades. The role of this systems is to ensure the protection of people, also the protection of main places such as airports, train stations and ports, securing the industrial and administrative buildings, to secure open area.This thesis presents a whole framework for detection and tracking in a camera network, the three main steps: detection, mono-camera tracking and people re-identification are addressed for multi-camera context. This study makes the following contributions: 1. Proposed a method based on an adaptive learning rate GMM for moving object detectionA dynamic learning rate GMM(Gaussian Mixture Model) algorithm is proposed to deal with the problem of slow adaption of GMM in the case of moving object detection in outdoor surveillance, especially in the presence of sudden light illumination changes. The GMM is mostly used for detecting objects in complex scenes for intelligent monitoring systems. To solve this problem, a mixture Gaussians model has been built for each pixel in the video frame, and according to the scene change from the frame difference, the learning rate of GMM can be dynamically adjusted. The experiments showed that the proposed method have a good results with adaptive GMM learning rate when we compared it with Gaussian Mixture Model method with a fixed learning rate. The method was tested on some dataset, tests in case of sudden natural light changing showed that our method has a better accuracy and less false alarm rate. 2. Proposed a method based on the trajectory of the moving object direction for scale changes problem in the mean-shift trackerObject tracking is an important part in surveillance systems, the one of the algorithms used for this task is the mean-shift algorithm because of its robustness, computational efficiency and ease of implementation, but the traditional Mean-shift cannot effectively track the moving object when the scale changes because of the fixed size of the tracking window and can lost the target while an occlusion, in this study a method based on the trajectory of the moving object direction is presented to deal with the problem of scale changes. Experimental results on some dataset showed that the improved method has a good adaptability to the scale. 3. Proposed a method for the occlusion problem in the tracking stepAfter the scale changes problem with the mean shift algorithm, we addressed the problem of occlusion, a histogram similarity metric is used to detect when target occlusion occurs in such situation a method based on multi kernel is proposed to estimate which part of target is not in occlusion and this part will be used to extrapolate the motion of the object and give an estimation of its position. Experimental results showed that the improved method has a good adaptability with the occlusion of the target while tracking, and also for the case of multi target tracking. 4. Proposed a multi-patch matching method for person re-identificationStarting from the assumption that: the appearance(clothes) of a person does not change during the time of passing in different cameras view’s field, which means the regions with the same color in target image will be identical while crossing cameras. First, we extracted distinctive features in the training procedure, where each image target is devised into small patches, the SIFT features and LAB color histograms are computed for each patch. Then we used the KNN approach to detect group of patches with high similarity in the target image, after that we used a bi-directional weighted group matching mechanism for the re-identification. Experiments on a challenging dataset(VIPe R) showed that the performances of the proposed method outperform several baselines and state of the art approaches.