Target Tracking related topics

1. occlusion

In short, there are three types of occlusion: Self-occlusion, inter-object occlusion, and background occlusion. Generally, it blocks objects and tracks multiple objects.Algorithm(For example, MacCormick and blke [2000] And Elgammal et al. [2002]) This problem can be solved based on the prior knowledge of the target location and appearance. However, partial occlusion of the scene structure is difficult to judge, because it is difficult to identify whether the target character changes or the occurrence of occlusion.

The general method to deal with occlusion is to use a linear or nonlinear dynamic model to model the target motion, and continue to predict the target position when the occlusion occurs until the target appears again and then correct it.

Beymer and konolige [1999] uses a linear velocity model and uses a Kalman filter to estimate the position and motion of the target.

Isard and MacCormick [2001] uses a nonlinear motion model and particle filter for state estimation.

Or use other features:

Contour projection-[haritaoglu et al. 2000]-positioning the character's head in partial occlusion;
Optical Flow features-[dockstader and tekalp 2001b]-assume that two targets are moving in the opposite direction;
Alternatively, it can be implicitly solved.

Sato and Aggarwal [2004] solves this problem with TSV

The freely formed target Contour Tracking operator uses different occlusion solutions. These algorithms usually use pre-established or online pre-built prior shapes to solve the occlusion problem.

Cremers et al. [2002] Build a shape model from possible target traits sub-space analysis to fill the section with lost outlines

Yilmaz et al. [2004] uses a Level Set contour expression-based hybrid model to establish an online prior shape. These methods can deal with complete object occlusion.

2 multi-camera tracking

Multi-camera tracking needs come from two aspects: one is the need for deep tracking information and the need for occlusion processing; the other is the need to expand the tracking scenario area.

An important issue of multi-camera tracking is the relationship between different camera perspectives.

Can be manually defined ([Collins et al. 2001; CAI and Aggarwal 1999])

It can also be automatically calculated based on the observed scenarios and objectives ([Lee et al. 2000; Khan and Shah 2003])

Another important issue is the complexity of dense computing.

Mittal and Davis [2003]-only sparse depth ing is calculated without dense Depth Estimation, reducing the computational workload.

Sports camera:

The system used by Kang et al. [2003] is a combination of a static camera and a pan-tilt-zoom camera.

There is no overlap between the camera angle of view: It is necessary to process sparse and discontinuous target observations. Make assumptions about the speed and path of the target. These algorithms depend heavily on the target's compliance with the hypothetical path and the expected interval between different cameras. If these conditions are not met, the target can only be used for any movement, that is, the single camera method.

This article from the csdn blog, reproduced please indicate the source: http://blog.csdn.net/lynphoenix/archive/2011/02/18/6192576.aspx