Research On Key Technologies Of Dynamic Objects Detection, Tracking And Recognition For Video Surveillance

Moving objects detection, tracking and recognition in video sequential images are the main research hotspots in the field of computer vision. Object detection and tracking are to accurately find objects from background images and to provide refined, accurate, invariant and differentiable features for object recognition and behavior analysis. Therefore, video object detection and tracking have become core technologies in video surveillance, machine vision system, visual precision guidance and many other application directions, and not only possess theoretical significance but also have broad application prospects.For the research in video object detecting and tracking, object segmentation and tracking algorithm based on the theory of active contour model is one of the fastest research development directions in recent years. Active contour model theory combines high-level prior knowledge and lower-level visual task, realizes the processing of visual task from top to bottom, and provides new theoretical framework for the solution of computer vision problem. In this thesis both theory and applications are considered for object detection, tracking and behavior recognition aspects. Main research contents and results in the thesis are the following:According to the disadvantages of the commonly used moving object detection methods, a new moving object detection method based on spatio-temporal union is proposed, which uses the idea of time segmentation first and then space segmentation. Firstly, the adaptive threshold is used to get difference image between the current frame and background image, and by analyzing the connected regions and setting a threshold, small noise are removed and moving object region are marked. Then, in the new method, boundary-based binary level set+Gaussian smooth filter is adopted to detect accurate object contours in extracted moving region. The binary level set is based on geometry active contour framework and the proposed binary level set function replaces the traditional level set function in order to avoid initializing the signed distance functions needed by the traditional level set function. Hence, the proposed method is not only simpler in implementation, but also needs a lot of less computational work. Moving objects detection of rigid or non-rigid is realized under the condition of static camera. The experiment result shows that the proposed approach is effective and feasible in real video environment, can obtain close and entire moving object contours.The GAC model is normally used for gray image segmentation. In this thesis it is applied to vector images. Hence a new spatio-temporal moving object detection method is proposed. The new method combines the improved variational GAC vector model based on level set theory with background subtraction. By considering the moving window obtained by time segment based on background subtraction as the initial contour of curve evolution and introducing the distance regularization term into GAC vector model, the need of level set re-initialization is eliminated, but also the numerical error is avoided. The regularity of the level set function is kept during the evolution, accurate moving object contour is obtained and the efficiency of the evolution and stability are improved. Experimental results show that the proposed method is applicable to both rigid and non-rigid objects, achieving good detection effect even in the case of partial occlusion.It is well-known that the geometric active contour model based on level set can better handle the variations of the curve topology. In order to track and get contour information of moving objects, a new linear moving object tracking method is proposed, which is the combination of the improved geodesic active contour (GAC) vector model with the Kalman filter. Firstly certain threshold is set, moving region with area either being too large or too small is deleted in the proposed method, and then for the rest of the objects GAC vector model is adopted to evolve curve in moving window regions, making the evolution curve approaching to the true contours of the object; Object features are extracted after obtaining object contour curve. By calculating the area of contour and centroid parameters, object and its relevant features are obtained (object contour area, object centroid). The object tracking is realized by using Kalman filter to predict the object position of the next frame.According to the degeneration phenomenon of particle filter, for nonlinear object tracking, a certain model cannot describe the change of the object contour in motion, thus increases the accuracy of tracking and extracting. In this thesis, inspired by the work of Tony Chan who studied object tracking based on region CV model, a new moving object tracking algorithm based on improved LBF model without re-initialization is proposed. Firstly, moving regions of object are obtained by using Gaussian particle filter, then the distance regularization term is introduced to improved model, furthermore the level set is initialized based on the centroid of external rectangle, and finally curve evolution is performed by using improved LBF model to get the precise object contour. In the end, the extracted result were fed back to the tracking frame, updating the reference template dynamically. An ideal tracking effect is obtained.Object behavior identification belongs to high-level vision in computer vision. The behavior is divided into two categories, namely,"events" and "activities". In this thesis, taking the objects appearing in traffic video as an example, we propose a novel traffic incident detection algorithm based on multiple features and Hidden Markov Model(HMM). Choosing detected events according to user needs, people often focus on the events such as collision and overtaking. Therefore we choose collision and overtaking as detected incidents. Using the vehicle detection and tracking algorithm, vehicle running trajectories are obtained, and then multiple features vector fusion method is proposed as an HMM input. For each pair of vehicles involved, we extract change of velocity of each vehicle and interaction feature as multiple features, traffic incident is modeled and recognized.