Research Of Long-Term Nonspecific Target Tracking Technology Based On Online Learning

Long-term nonspecific target tracking technology refers to continuous effectively tracking of interested target area in a video. A key problem is, for the videos in which the target motion is intense, how to track the object after the shape has changed. Online learning during the target tracking process is one of the most useful methods to deal with it.This paper gives a kind of target template update method based on online learning, which is, to determine the positive and negative samples and update the target template by calculating the overlap between the scanning windows and the target area, the overlap is defined as the ratio of the intersection of the two rectangular boxes to the mean area. Then, based on Tracking-Learning-Detection (TLD), using the target template update method of given, this paper carries on the experiment analysis. At the same time, the TLD tracking system is improved, which can achieve reliable and accurate target tracking, and improve the tracking success rate of the system.The main work of this paper is:1. A new method of target template update based on online learning is proposed, and the method is applied into the TLD tracking system to improve the tracking success rate of the system.2. Through the research on the initialization of the TLD tracking system, the selection method of the tracking target is improved. Changing the mouse to drag the formation of a rectangular box as the initial target into the mouse click to select the rectangular center, the formation of the initial target. Through which reduced the interference of the background area and made the operation more convenient.3. Improved the detection module of TLD, quoted frame difference method before the variance filter, through which improved the accuracy and efficiency of the detection module.4. On the VS2008 and OpenCV2.3.1 platform, this paper compares the original method and the improved method.For several groups of representative video sequences, the performance of the improved system is better than before. Among them, in the video sequences whose target motion is more intense, the tracking effect is particularly significant, that is, the improved system can be more adapted to the deformation of the target, and can actualize long-term tracking.