A Research On Video Tracking Based On Online Learning

Video tracking is one of the most important research direction in the field of computer vision, it combines research and advanced technology in the field of pattern recognition, image processing, image characterization and computer applications and has a very important research value and practical significance in intelligent monitoring traffic control, human interaction and industrial control. We did some useful research and exploration of video tracking algorithm, and conducted in-depth research in long term tracking for video and adaptive recovery issues. The major work are as follows:1:Detail study in the tracking algorithm based on the KLT feature point,and to the situation of quickly tracking failure, we extend it to the multi-scale image space, construct a Gaussian image pyramid to tracking the feature point. Meanwhile, introduce a new error forward-backward error to filter the feature points, the accuracy of the feature point is greatly improving the tracking algorithm, and finally use the Greedy algorithm to converges the target position, the experimental results show that this method is effective.2:Treat the tracking as a binary classification problem and proposed an incremental random forest classifier support online learning, it can effective deal with the problem of small sample data streams such as video tracking. Construct classifiers online, and update the classifier online using the samples obtained by the sequential samples, so as to adapt to the changing in the process of target motion, and to prune the classifier tree structure based on these changes. The quantitative statistics show that the classification accuracy of the video tracking and recall rate with good accuracy and stability in a complex environment.3:Aim at the long target tracking and the adaptive recovery, we proposed a new tracking method based on the LK flow and online learning. The algorithm based on the weight displacement scan method to weight extract samples, and use the set of constraints on target motion changes. use the key-frames to build observation model and motion model to update the ensemble classifie,. to make the classifier online learning target’s shape changes. After the failure of the target tracking, the detection module to re-detect the target for re-initialization of the short-time tracking algorithm in complex background, moving target scale changes in the appearance of changes in the scope, shelter and disappearance cases remains stable tracking target, and can be automatically recover after the tracking failure.