Transferred from: http://blog.csdn.net/wjbwjbwjbwjb/article/details/7169220
Recently read an article on Tpami, which is about multi-instance learning tracking.
Very appreciative of its article writing, the introduction summarizes three basic elements of the tracking system composition: The appearance model, the motion model and the search strategy. The author's work mainly aims at the appearance model, constructs a kind of multi-instance online learning method, unifies the Milboost and the online AdaBoost, the effect is also very good.
But I'm talking about tracking innovation, so here's a little bit of my understanding:
1) Appearance model. That is to study how to better evaluate the similarity likelihood problem of a candidate target location, such as: detection-based tracking, multi-instance learning tracking, incremental learning tracking, and so on. The basic idea is to improve the evaluation result of target similarity and improve the adaptive ability of tracking. The main method comes from the development of detection and recognition technology, which will be used in the tracking of the excellent methods of detection and recognition in recent years.
2) motion model. This is a control problem, the current research is relatively few, compared to classic such as Kalman filter method, autoregressive sliding average, particle filter in the first order, Ishimarkov process, want to break through in this aspect may need in cybernetics, automation and other methods have certain research, put forward a better motion model to depict the movement of objects.
3) Search strategy. In fact, the most important thing about this problem is the mean shift algorithm (Mean Shift), which is a very classical local optimization search method, and also the most significant application of the steepest descent method in visual tracking. In addition, the research on the optimization of the comrades should be able to make relatively good innovation.
In view of the above analysis, I am doing a study on the third issue.
The breakthrough of visual target tracking innovation point--a brief discussion (turn)