Research Of Key Technologies Of Pedestrian Tracking Across Non-overlapping Camera Views

Pedestrian tracking across non-overlapping cameras refers to detecting, tracking and reidentifying the pedestrians between different cameras in the non-overlapping view by using computer vision, pattern recognition and machine learning techniques. The purpose is continuously tracking the pedestrians for a long time in a wide area. Currently, it has become a research focus as well as one of difficulties of intelligent transportation system. However, pedestrian tracking across non-overlapping cameras still faces enormous difficulties and challenges in the actual traffic monitoring environment, due to the presence of complex background, illumination variation, partial occlusion, motion deformation, the camera shooting angle and other factors, such as interference between the cameras from their own property differences. For the disturbance factors above, this thesis is trying to tackle the problem of pedestrian tracking across non-overlapping cameras in aspect of detecting, tracking and reidentification. The main contributions include:1. This thesis proposes a significant binary Haar-like feature(SLBH), which adds description of significant pedestrian information based on integrating merits of the LBP features and Haar-like features. In addition, for the reason that pedestrian detection algorithm based on the overall feature is easily affected by partial occlusion and motion deformation, this thesis proposes a two-stages pedestrian detection algorithm based on multi-component validation. Firstly, the algorithm adopts low-dimensional SLBH feature to detect the entire pedestrians, and secondly uses body parts for further testing of candidate pedestrians falling into the gray zone, and finally applies Bayesian algorithm to estimate and analyze candidate pedestrians in the fuzzy zone. Theoretical analysis and experimental results show that the proposed algorithm can effectively reduce the false alarm rate of pedestrian detection.2. The thesis proposes a pedestrian tracking algorithm with adaptive feature fusion to automatically adapt to changes in the external environment in the particle filter framework. At the same time, in order to reduce the time cost overhead of particle feature extraction generated from multi-feature fusion, the thesis proposes a method similar to the integral image technology to accelerate extraction speed of particles feature. Theoretical analysis and experimental results show that the proposed algorithm can effectively adapt to changes in the surrounding environment and improve the accuracy of pedestrian tracking.3. The main work of pedestrian re-identification is to confirm the same pedestrian between two cameras. In the pedestrians walking process, it is prone to motion deformation, partial occlusion and illumination variation. Besides, differences in their own property of shooting angle of the camera, exposure time and other factors have a direct impact on the accuracy of the pedestrian re-identification. Therefore, in order to increase the robustness of pedestrian reidentification algorithm for factors above, the thesis proposes a pedestrian re-identification algorithm based on group similarity comparison model. Firstly, the algorithm constructs pedestrian sample prototype in the gallery and takes advantage of sample prototype similarity feature instead of low-level features to increase the semantic information of features. Secondly, the algorithm applies an image sequence instead of a single image as a probe, and employs systematic sampling method to group the image sequence, and calculates similarity for intragroup sample images while not calculating similarity for extra-group sample images to construct individual differences similarity feature. Finally, features are fused using Ada Boost classifier to implement pedestrian re-identification. In addition, in order to enhance the ability of the Euclidean distance to distinguish similarity measure between two feature vector, this paper proposes a significant difference distance. Theoretical analysis and experimental results show that the proposed algorithm can effectively resisting various affine transformation and noise in pedestrian traffic scene.4. In order to enhance the reliability of the information in the non-overlapping camera network in spatial and temporal clues, the thesis proposes the cameras network topology estimation algorithm based on weighted time. The algorithm assumes that the transition time between different cameras of pedestrians could meet the Gaussian distribution, and by giving greater weight for pedestrians whose arrive time approaching to the average transfer time, while giving less weight value for pedestrians whose arrive time far away from the average transfer time, it highlights contribution for estimating network cameras topology by the different time of the pedestrian arriving. On this basis, combined with pedestrian detection, pedestrian tracking and pedestrian re-identification algorithm, the thesis proposes a spatial and temporal fusion pedestrian across non-overlapping camera tracking algorithm. Theoretical analysis and experimental results show that the proposed algorithm can effectively improve the camera network topology estimation accuracy, and can be better accessibility to pedestrian tracking across non-overlapping cameras.