Research On Visual System Based Pedestrian Detection And Tracking Method

Computer vision-based pedestrian detection and tracking has increasingly become one of the most active research topics in the fields of pattern recognition, computer vision and artificial intelligence, which has important applications in the fields of intelligent video surveillance, intelligent transportation, advanced human-computer interaction, abnormal behavior analysis and virtual reality. The core of computer vision-based pedestrian detection and tracking is that using the theoretical achievements of pattern recognition to detect and track pedestrians by combining computer vision technology from a video sequence, the research topic involves multiple theoretical disciplines such as image processing, pattern recognition, system state estimation, probability and statistical inference. Meanwhile, computer vision-based pedestrian detection and tracking, as an underlying technology, is the basis of the higher level behavior analysis and understanding in the field of computer vision, its detection and tracking accuracy will directly affect the subsequent varieties of advanced applications, such as target recognition, motion analysis, and pedestrian behavior understanding etc.For machine learning and statistical classification based pedestrian detection algorithm, there are two key issues closely related with the detection results, the first one is the design and extracting of pedestrians features, and the second is the choice and design of the classifier. According to these issues mentioned above, the paper mainly studies pedestrian detection algorithm based on machine learning and statistical classification. The detection method lies in the description to the object appearance and shape, effective feature extraction and efficient classifier design. With regard to the design and extraction of pedestrian features, according to the overall consistency characteristics of pedestrian appearance or shape, and the pedestrian saliency features compared with other object category, such as bilateral symmetry, strong vertical edges presented along the boundaries of the body, this thesis proposes three kinds of pedestrian features, relational color similarity feature, random color similarity feature, saliency based binary feature. With regard to the choice of classifiers, the thesis employs Semi-Naive Bayesian based random Fern classification to classify the candidate sample, and achieves promising classification performance. In addition, the real-time problem of pedestrian detection is also needed to be considered, for which, the issue of how to improve pedestrian detection speed in infrared scene is also discussed.In complex scenes, visual tracking usually turns into a non-Gaussian/nonlinear state estimation problem. One of the powerful tools for such non-Gaussian/nonlinear state estimation problem is particle filter, which also is a complete theoretical framework for constructing visual tracking algorithms. At the same time, multiple visual feature fusion tracking can be easily realized by combining observation probability in the particle filter theoretical framework. And then, the thesis explores a particle swarm optimization based multi-visual features adaptive weight particle filter tracking algorithm. The another challenge to track pedestrian target is that the human body is a complex nonrigid structure, which will generate larger attitude change during the movement, for which, the thesis proposes a feature block and sparse representation based pedestrian target tracking algorithm. The proposed tracking algorithm can achieve fine tracking of pedestrian object, I.e. Given that pedestrian objetc is considered as an entire tracking object, the proposed algorithm can have a good tracking performance, and furthermore, the proposed method can also effectively track the limb movement of pedestrian object.How to improve the vision-based pedestrian detection accuracy and pedestrian object tracking robustness is the major research content in the thesis, with regard to the specific problem in the pedestrian detection and tracking process, effective solution algorithms are proposed, the creative points in this thesis contain:1. As a low-level feature, color feature is very popular in the computer vision research fields, in order to use the low-level feature owning abundant amount of information, an RGB color space based pedestrian feature descriptor is proposed in the thesis, which is also called relational color similarity feature(RCS). The feature descriptor has no direct relationship with the color value, but can well describe the pedestrian inherent bilateral symmetry property and upright property.2. Naive Bayesian based random Fern classifier owns desiring classification performance and fast classification speed, but the corresponding training or classification sample feature needs to be binary feature, in order to utilize the classifier, the thesis explores a novel color based feature called random color similarity feature. Due to constraints on the manufacturing of printed cloth, clothing textures are different from natural textures, which seem to be significant in the human visual system, especially the contact edge between pedestrians and the corresponding background, based on which, saliency detection based binary feature is proposed in the thesis.3. In infrared scenes, the pixel by pixel traversal detection method usually leads to low detection speed. In order to improve the pedestrian detection speed in infrared scenes, saliency detection based pretreatment strategy is presented, the image region possibly involving pedestrian is firstly extracted by the pretreatment strategy, and then, pedestrian detection is conducted in such image region, which can greatly increase the pedestrian detection speed in a certain extent.4. In practical applications, since it is difficult to achieve robust tracking by utilizing a single object feature, particle swarm optimization based multi-cues weight adaptive particle filter tracking algorithm is proposed in the thesis, experiments performed on several challenging pedestrian sequences demonstrate the effectiveness of the proposed tracking algorithm.5. Aiming to the posture diversity of the non-rigid pedestrian, sparse representation based image blocks tracking method is proposed. Since image sparse representation is not sensitive to noise and occlusion, the adverse effects caused by the background noise and occlusion is weakened by using image sparse representation as the observation model. At the same time, an on-line template update strategy which uses the incremental subspace learning method is employed to adapt to the object appearance changes caused by the illumination or pedestrian posture changes. The experiment results show that the proposed approach achieves robust tracking even when the pedestrian appearance and posture have large changes.