Intelligent Video Surveillance, Moving Target Tracking

Intelligent video surveillance is an important research topic in computer vision. Its main task is to detect, classify and track the targets of interest from the video images, and analyze, understand, even describe their behaviors. Target tracking technique plays a key role in intelligent video surveillance. In order to solve the target tracking problems in different scenes, this thesis focuses on three target tracking algorithms, including Kalman Filter, Mean Shift and Particle Filter. The innovations and the main work of this thesis are as follows:1. Kalman Filter is combined with SVM (Support Vector Machine) to track the two targets (person and car). Firstly, the targets are detected and then SVM is used to classify and recognize them. In this way, the corresponding relationship between the centroids and the targets is figured out. Then each centroid is sent to the corresponding Kalman Filter. The person and the car can be tracked correctly with the two Kalman Filters.2. Develop the target tracking algorithm combining Mean Shift, which is based on color cue and texture cue, with Kalman Filter. Firstly, the Mean Shift algorithm based on color cue and texture cue is used to calculate an accurate location in the current frame. The coordinate of the location is sent to Kalman Filter as the measurement. Then Kalman Filter is applied to predict the next initial searching location for the Mean Shift iterations in the next frame. With this method, the target can be tracked successfully even when there is a large movement between two consecutively processed frames. Besides, this algorithm is also effective in the environment where the color distribution is similar between the target and the background. In such an environment, the target can not be tracked correctly with the Mean Shift algorithm based on the histogram in the RGB color space. Experimental results show that this algorithm is effective, robust and can be used for tracking in different scenes.3. For the target tracking problem in the nonlinear and non-Gaussian environment, the Particle Filter algorithm based on color cue, texture cue and edge cue is proposed. The tracking robustness is improved since more cues are integrated into the Particle Filter. Firstly, process the first frame after the initialization. Compute the color PDF (Probability Density Function) of the target by the histogram in the HSV color space. At the same time, the texture cue and the edge cue are calculated for modeling. Secondly, process the next frame. Finish the process of state transition. Thirdly, calculate the likelihood model of the multi-cue integration. Then compute and normalize the new weights of the particles. Finally, estimate the state of the target according to the MMSE (Minimum Mean Square Error) criterion. When the effective size of the particles is below a certain threshold, the resampling procedure will be done. The experiments include the target tracking in the complex background, the target tracking in the environment where the short-term partial occlusion and the short-term complete occlusion occur, and the comparision of tracking results between the Particle Filter algorithm and the Mean Shift algorithm. The experimental results show that the proposed algorithm is effective and robust.