Study On Video-based Detection Andtracking Method Of Pedestrian In ITS

Visual pedestrian detection and tracking is a key problem in computer vision and ITS, and it is the basis of follow-up treatment, such as visual scene analysis and semantic analysis. The technique of visual pedestrian detection and tracking has a wide range of applications in intelligent video surveillance, traffic, human-computer interaction, visual navigation of robots, virtual reality, medical image processing, national defense, etc. Although the pedestrian detection and tracking has been studied for more than ten years in computer vision community, it is still an active research area. In present, there is no pedestrian detection and tracking system which is general, robust, accurate, efficient, and real time. Because the human body is non-rigid, the scenes are cluttered, and there are a lot of interactions among the pedestrians or between the pedestrians and the scenes, the human detection and tracking is one of the most difficult challenges in the area of computer vision research.In this dissertation, some key issues of pedestrian detection and tracking have been discussed, including accurate segmentation of background, objects and shadows, accurate pedestrian recognition, faster speed of tracking, more accurate location and pedestrian's activity understanding and description. The highlights and main contributions of the dissertation include:(1) Accurate segmentation of background and objectsThis paper mainly does the research on moving object detection in video sequences. To point against the shortcomings of the existing moving object detection algorithms, this paper proposed a kind of more robust and better qualified to real time request object detection algorithm, and section-distribution background model. The object detection algorithm includes pre-process, section-distribution modeling, denoising, foreground extraction and background update and so on. As the basis of this approach, the section distribution model aimed to establish a rapid, accurate, and strong adaptive background model. In order to better adapt the changes brought by illumination, weather or other factors better, the algorithm contains timely background update strategy.(2) Accurate segmentation of shadowsA new shadow detection algorithm based on the Gabor wavelet and the color model is proposed in this paper. Firstly, the Gaussian mixture distribution model and shadow color model of the background are established. Next, the foreground figures are extracted by means of the difference method. Then, the potential shadow pixels are found out by means of Gabor wavelet texture analysis, which are further analyzed with the shadow color model to search real shadow pixels. Finally, the real shadow areas are distinguished.(3) Pedestrian detection under static backgroundThis paper does a research on pedestrian detection issue in video, and presentes a pedestrian detection approach based on triangle feature set, which adopted the proposed triangle feature set. Firstly, this paper uses the rectangle feature for reference to describe the edge feature, through analyzing the edge feature of the pedestrian posture to obtain new feature set—triangle feature set. Secondly by mixed rectangle feature, triangle feature and asymmetrical feature together, this paper raises a new feature set - Hybrid feature set. Meanwhile, to point against the conventional Adaboost algorithm (Adaboost algorithm based on rectangle feature set) which exists overfitting, an improved Adaboost algorithm has been presented. The algorithm utilized the hybrid feature set, and optimized the threshold selection strategy, the weight update strategy and the normalized process,and improved the sample training procedure of the primary algorithm. At last, considering the feature that the size of pedestrian in video would change with the different distance between the camera and the pedestrian, this paper presentes a multi-scale window traverse strategy.(4) Pedestrian detection against camera shiftStudy on pedestrian detection against camera shift, this paper proposes proposed an optimization pedestrian detection algorithm based on quantum evolution. This approach bases on AdaBoost pedestrian detection algorithm, supporting vector machine (SVM) and multi-objective optimization theory as the basis, and the core of the approach is quantum evolution which bases on real encoding. Firstly, it utilizes the AdaBoost to classify pedestrian with coarse granularity, and then employ SVM to design more accurate pedestrian detector. Taking multi-parameter with complex relationships and no reasonable regulation criteria into account, this paper considers the construction condition of kernel function, introduce real quantum evolution algorithm to the domain of SVM parameter optimization problems, and adopts multi-objective optimization concept to enhance the classification performance, which achieves good results. Meanwhile the complexity of the algorithm has been analyzed in theory to ensure the real-time characteristic.(5) Pedestrian trackingThis paper proposes a pedestrian tracking algorithm by using improved CamShift algorithm and incidence matrix. Firstly, a position algorithm is presented by an improved CamShift, which has been used to make accurate position on pedestrian object. To solve the tracking problems under complex scenarios, including the convergence, occlusion, sudden disappear and appear of the pedestrian objects and so on, this paper proposes an algorithm based on multi-factor incidence matrix. Firstly, section distribution model and shadow detection has been used to acquire foreground blob. Secondly, prediction of object position has been realized by using Kalman filter. Thirdly, the improved CamShift algorithm has been used to realize accurate position. Fourthly, the similarity between the foreground blob and the predicted object has been analyzed as well. Then the multi-factor incidence matrix between foreground blob and object has been setup. Finally the pedestrian accurate tracking under complex scenarios has been completed by judging and reasoning according the incidence matrix.