Discussion on target tracking development direction

Current Target TrackingAlgorithmRobust real-time tracking in simple scenarios. However, some constraints are usually required, such as the smoothness of motion, the number of blocks, the constant illumination conditions, and the contrast between the target and the background. It is difficult to meet these conditions in actual application scenarios. Therefore, problems related to feature selection, target expression, motion shape, and motion estimation are still quite popular research fields.

Tracking targets in unrestricted daily videos is also a major challenge. Noise, compression, and unstructured scenarios are all difficult factors.

Using specific environment information is also an important research direction. For example, cars in vehicle tracking should be restricted on the road rather than on the sky or walls. This information is very meaningful for identification.

The selection of feature sets has a great impact on the tracking effect. Generally, features with better differentiation are selected.

Nowadays, weighted combinations of multiple features are widely used. Machine Learning and pattern recognition have done a lot of research on feature selection, but both require offline feature learning. However, tracking problems requires online feature selection. [Collins and Liu 2003; Stern and Efros 2002] Some research has been conducted on this issue, but further research is required. Online boosting [oza 2002] may be a suitable solution.

The choice of the target expression model is also an important part. At present, most algorithms use predefined models, while online model learning can improve tracking performance.

Motion-based segmentation [Vidal andma 2004; black and Anandan 1996; Wang and Adelson 1994] and multi-objective factorization methods [costeira and kanade 1998; gear 1998] is used to learn the moving model of multiple targets in scenarios. However, these methods assume the rigid motion of the target. Unsupervised learning of the Target Model in non-rigid multi-object motion is still difficult to solve.

Semi-supervised learning modeling may provide better solutions. Including cotrainging [Levin et al. 2003; Blum and Mitchell 1998], transductive SVMs [joachims 1999], constrained graph cuts [Yu and Shi 2004]; these methods do not require a large amount of training data, ability to learn non-rigid target shapes and appearances, and may also include background information in the form of negative training data.
Probability state space methods, including Kalman filtering, jpdaf, hmm, and dynamic Bayesian networks, are used to estimate target motion parameters. The Dynamic Bayesian network is the most common method to express the conditional dependence between multiple variables or image observations. It also provides a standard framework for information fusion from different information sources. However, the accuracy of these algorithms needs to be further improved.

This article from the csdn blog, reproduced please indicate the source: http://blog.csdn.net/lynphoenix/archive/2011/02/18/6192592.aspx