Online video target tracking method based on depth cross similarity matching

Abstract

The invention belongs to the technical field of video target tracking, and provides an online video target tracking method based on depth cross similarity matching. The method comprises the followingsteps of designing a depth feature cross similarity module, capturing all local similarity information of a template and a sample, wherein obtained similar features are no longer sensitive to displacement and deformation; designing a similarity attention layer, distributing weight coefficients for cross similarity results of different spatial positions, and enabling a tracking algorithm not to respond to edge background interference; and designing a loss function containing a parameter regularization item, and rapidly optimizing the parameter to an optimal value. Based on the above three-pointbasic scheme, the deep learning twin network based on matching is used as a basic framework, and any target in the video is tracked in a precise and robust manner. From the tracking effect, the method provided by the invention has the capabilities of distinguishing similar objects, re-identifying a reappearing target after shielding and coping with rotation and deformation, and can be applied tovideo applications such as automatic driving of a front scene, autonomous flight of an unmanned aerial vehicle, traffic or a safety monitoring video and the like.

Description

technical field [0001] The invention belongs to the technical field of video target tracking. Given any target object specified in the first frame, the position and scale of the target can be framed in subsequent video sequences, and involves related theories of digital image processing, deep learning and linear algebra. Background technique [0002] Thanks to the wide application of artificial intelligence technology and the maturity of deep learning methods, the video target tracking algorithm, as the core basic technology, can be widely used in video surveillance, UAV aerial photography, automatic driving and other fields, and has a certain research basis. [0003] There are currently two mainstream directions for video target tracking methods: correlation filtering methods and deep learning methods. The correlation filtering method is based on a cyclic sampling strategy, and continuously optimizes and updates the filter for the prediction of the next frame by modeling t...

AI Technical Summary

Problems solved by technology

[0006] The technical problems to be solved by the present invention are: first, given any scene and specifying any target therein, even if the appearance of the target changes (such as non-rigid target deformation, turning from the front to the back, etc.) After reproduction, the background noise interference is serious and the situation that is difficult to distinguish from similar appearance objects, the present invention still maintains robust tracking; second, the method of the present invention has portability, and the solution tracking method exists independently and is not easy to be compatible with other related video understanding tasks ( Such as pedestrian search, target detection) integrated multi-task technical solutions

Images