Mean Shift-based Video Object Tracking, Detection Algorithms And System Implementation

Video target tracking is an important component in the computer vision field, which combines pattern recognition, artificial intelligence, automatic control and many other related fields of knowledge. Video target tracking, extracted the target location information from image sequences, can be used for intelligent video surveillance, video-based human-computer interaction, automatic driving, agriculture automation and medical imaging and other fields. In recent years, Mean Shift algorithm, with good real-time, anti part of the block, not very sensitive to deformation and good robustness has aroused great concern in video target tracking application. However, it also has some shortcomings, such as the window width can not be adaptive, and it can not effectively track rapid and large-scale shelter objectives, and it need to initialize the tracking object manually. Mean Shift algorithm has two versions: the standard Mean Shift algorithm and the CamShift algorithm. This thesis based on the Mean Shift algorithm in video object tracking applications; targets these shortcomings to make good improvement. The main content and innovative as follows:(1)A detailed study of the Mean Shift algorithm in the field of video object tracking application, a better algorithm to solve the problem of adaptive window width.(2)Presents a histogram based on the gradient tensor Mean Shift algorithm. The traditional Mean Shift algorithm use the color space feature model, when the target and background is similar in color, it will easily track failed. However, the goal of spatial texture information at this time there may be a good distinction between object and background. In thesis, the tensor gradient histogram method not only uses the objective texture information, but also target from the three-dimensional texture map directly to a one-dimensional space, reducing the computational amount and improving the target and the background color of the tracking results under similar conditions.(3)To track fast-target and block-target, thesis uses Kalman filter to predict the state in the next frame, and the predictive value of the mean shift algorithm is used to the starting point of search target. This thesis presents the use of similarity coefficient or Kalman residuals compared with the threshold value, when the similarity coefficient or Kalman residual is less than a certain threshold that the targets are obscured, only use Kalman prediction algorithm to track until the goal of re-emergence.(4)The use of AdaBoost target detection, target can be in advance off-line learning, then the characteristics of target can be achieved in the video sequence, so that the system can automatically detect the target and track the target automatically ,which solve the initialization problem.(5)In OpenCV framework, this thesis designs and completes an experimental video target tracking software system using modular implementation, including target detection, blob detection, blob tracking, tracking processing and trajectory generation of modules, which has carried out experiments verification, and achieved good results. The system for future research work provides a convenient experimental test.This thesis studies the static and motion video camera scenarios target tracking. The major research area of video target tracking is single-target tracking and part of block processing. Experimental results show that the algorithms used in real-time performance is good, strong anti-interference ability, and can be in more complex backdrop for the work.