Research On The Algorithm Of Target Tracking Based Online Learning

Moving target tracking technique is to detect moving targets and tracking the target. In video target analysis system, the technology provides a basis for further measures, such as behavior analysis, human-computer interaction, intelligent monitoring and so on. In complex tracking environment, target face many interference factors, such as posture,illumination changes, shade and motion blur, etc. An algorithm to for long time effective tracking of moving target is very difficult.In order to solve the problem that traditional algorithm in the complex environment cannot tracking the target for a long time. This paper proposes a tracking algorithm based on online learning. The core of the target tracking algorithm is LK(Lucas- Kanade) optical flow algorithms. In the process of tracking, LK algorithms of target appearance model needs to be update online which makes the model can transform to adapt to the target and environment. Therefore, the algorithm improves the accuracy of target tracking.The main research works in this thesis are as follows:1. An online updating background model and a motion trajectory model are established,and two methods are presented for judgment of whether the tracking target is lost. To determine trace whether the object is the goal, to avoid the tracking loss arising from background interference or targets in the images disappeared due to the tracking failure2. The working effect of algorithm depends on the accuracy of the template matching.If the template of the pixels is all pixels of the target, so LK algorithm has a good tracking effect. But if the template contains background pixels, may be affect the similarity of the template, make the algorithm drift. Algorithm for this, this article puts forward a new kind of template matching method, in the structure of template matching the introduction of the target information; improve the accuracy of the template matching.3. The traditional LK optical flow algorithm through the target model in the neighborhood of the best match to implement tracking of moving targets. And the best match is through solving the multi-scale iterative operation, easy to fall into local minimum value in the iterative computation. To this end, the paper put forward based on the Bayesian model derived from the pixel template matching method, the maximum likelihood estimates of the model for two-dimensional transformation to achieve the target tracking.