Multi-camera Non-rigid Object Detection And Spatial Localization System

With the popularization of video surveillance system, computer vision is becoming increasingly familiar to the public. As one of the most important parts of the field of computing vision, object detection and localization which have been broadly applied in fields like security monitoring system, intrusion detection, driverless cars, has attracted increasing concentrations. Traditional video surveillance system cannot adapt to complex scene because of its dependence of human participation. Intelligent surveillance can cover lots of scenes at the same time, analyses abnormal behavior and provide real-time alerts, while need seldom human work. This thesis mainly focuses on the study of object detection, applied several object detection algorithms to cameras and constructed an object localization system. Main works of this thesis include:1. A traditional object detection algorithm based on frame difference and connectivity checking is proposed. Target contour calculated by morphological process on difference image obtained by frame difference. Some strategies are used to address the issue of object fragmentation caused by non-rigid deformation. The method can get a rough location of the target in a very short time.2. An object detection algorithm used Deformable Part Model(DPM) based on statistical learning is proposed. The method used HOG descriptor as its model feature, which makes the model invariant to geometric and optical transform. As a result of deformable model, this method is robust to non-rigid deformation. The method works well especially in the field of pedestrian detection. Support Vector Machine and stochastic gradient descent are used to train the model to get a fast convergence rate.3. A multi-camera localization algorithm based on joint probability distribution is proposed. This method performs more accurate results than traditional methods which use geometrical center. An analysis of theoretical error of both the vertical and horizontal directions is performed to obtain the theoretical precision of the whole system.