Intelligent Visual Surveillance, Moving Target Detection And Tracking

Intelligent visual surveillance is one of the hotspots in the field of application for computer vision, which can locate, track and identify changes in the scene through automated analysis of video camera recorded image sequences without manual intervention. Moving targets detection and tracking are basis for intelligent visual surveillance.Our objective is to develop an algorithm for real-time Multi-target tracking (MTT) and to implement it on a DSP platform ultimately. Firstly, two commonly used algorithms for moving targets detection and tracking, Nearest Neighbor (NN) and Mean-shift, are studied comparatively. Single Gaussian method is chosen for background modeling. NN and Mean-shift are tested by PETS 2001 dataset. Results demonstrate NN's arithmetic speed is faster than Mean-shift, but tracking will fail if the tracking targets are too close. Also, since color feature is involved in Mean-shift, the operations will be slower and tracking will be much more accurate. Mean-shift is light-sensitive and its high computational complexity is great challenge for real-time tracking on DSP platform.A texture based MTT algorithm is developed in the dissertation. It chooses texture instead of color because in an outdoor environment background is complex. The useful texture descriptor, local binary pattern (LBP), is chosen since DSP is not good at floating-point operations. On the basis of NN, a new parameter is used to adjust the search window so that blobs which move too fast can also be found. The Kalman filter is introduced into the algorithm to predict the blob's new position and size. First of all, blobs are searched in the neighbor based on the Kalman predictions and the blobs which satisfy certain conditions are added into a candidate sequence. If just one blob is found, the blob is considered as the tracking target. In the case of the existence of multi-blob,χ2 LBP distance is calculated between the blob and its candidates. The candidate, which has the minχ2, is set to be the tracking target. LBP-based feature methods have low computational complexity, so the proposed method can be implemented on a DSP platform efficiently. Experimental results indicate that LBP is feasible for blob distinguishing. Due to the application of texture, the algorithm can cope with slow illumination changes and slight leaf-shaking.In order to track specific targets, an affine-model-based camera calibration algorithm is proposed at last. With target's pixel height in the image, the algorithm can calculate its true height in the real world roughly (inaccuracy≤20%), which can be applied to filter most part of targets that we are not interested in.Tracking results demonstrate the effectiveness of the algorithm. This algorithm has been implemented on PC and DSP platforms and achieved real-time performance.