Multitask learning object tracking method based on subspace characteristics

Abstract

The invention relates to a multitask learning object tracking method based on subspace characteristics. The method makes the most of the advantages of a subspace and sparse reconstruction at the aspect of target appearance modeling, enables an l1 algorithm to be introduced into the subspace, and employs a multitask learning (MTT) method to find the mutual relation among all particles. During the construction of a dictionary, an image template is not employed, and the characteristic subspace of a PCA is used for the construction of the dictionary. Moreover, a trivial template is added to reconstruct noise. The solution to the sparse expression coefficient of each particle is taken as one single-task learning problem in MTT, and the parse expression coefficients of all particles are solved through the application of the combination of normally-used mixed norms l2, 1. Moreover, a Accelerated Proximal Gradient (APG) method is used for the solving of the multi-task parse expression. Compared with a method of l1 tracking, the MTT method improves tracking effect and reduces the calculation complexity through the mining of the correlation among the particles.

Description

technical field [0001] The invention belongs to computer data image processing technology, in particular to a multi-task learning target tracking method based on subspace features. Background technique [0002] Object tracking algorithms have broad application scenarios. It plays an important role in intelligent video surveillance, intelligent transportation, robot vision, video indexing and other fields. In addition, object tracking has been widely researched and applied in the fields of artificial intelligence, content retrieval, precision guidance, human-computer interaction, and medical diagnosis. [0003] Although the research on object tracking algorithms has been carried out for many years, there are still many difficulties and challenges in this field. Object tracking is still one of the hot issues in the field of computer vision. At present, there is still no good algorithm that can comprehensively deal with the problems in the field of object tracking. [0004]...

AI Technical Summary

Problems solved by technology

Recently, some scholars have proposed some accelerated l 1 The method of optimization solution is applied to target tracking, but these methods all assume that the sparse representation process between each particle is independent of each other, ignoring the relationship between particles will easily lead to the drift of the target, especially when the target has a different appearance. When there is a significant change

In addition, these algorithms mainly use image templates as dictionaries, so that the over-complete dictionary has a high dimensionality and a large amount of calculation, and the original grayscale features of the image are easily affected by illumination and similar backgrounds.

Images