Multi-scene target tracking method based on adaptive depth feature filter

Abstract

The invention discloses a multi-scene target tracking method based on an adaptive depth feature filter. The method comprises the following steps of converting a target area of an original image into acolor naming space from an RGB space to reduce the interference of color changes, then, calculating a foreground probability graph of a target area, t and training by using the features extracted from the foreground area according to the foreground probability graph to alleviate the boundary effect and effectively inhibit the noise in the background, so that the method can adaptively extract thetarget features. According to the method, multiple layers of depth features are used for training in corresponding related filters respectively; and the weight of the corresponding depth feature filter is updated adaptively according to the tracking effect, the filter stability, the historical response and other information, and the tracking model is guided to adaptively select useful depth features in different scenes, so that the target can be tracked in a robust manner in various complex scenes. Compared with the prior art, the method has the advantage of high robustness, and can accurately track the target in various complex scenes.

Description

technical field [0001] The invention belongs to the technical field of computer vision, relates to video target tracking in multimedia technology, and is a multi-scene target tracking method based on an adaptive depth feature filter. Robust tracking of objects in the scene. Background technique [0002] The task of video target tracking is to use the marked target area information in the first frame to automatically predict the target area information in subsequent frames, which includes the position and size of the target area. Video object tracking is one of the areas that researchers focus on, and has already achieved many practical applications in real life, such as eye tracking, automatic driving, and intelligent monitoring. In general, according to different target tracking models, target tracking algorithms can be roughly divided into generative model-based tracking and discriminative model-based tracking. Generally speaking, a typical generative model-based object ...

AI Technical Summary

Problems solved by technology

[0004] The problem to be solved by the present invention is: the boundary effect existing in the existing video target tracking technology cannot well integrate the depth feature, and the depth feature filter weight is fixed, The learning rate of the model is fixed and cannot adapt to a variety of complex scenarios.

Images