A target tracking method based on online learning

Abstract

The invention discloses a target tracking method based on online learning. The method comprises the following steps: initializing the target, initializing the basic information of the target by tracking the target position and size given in the first frame of the video; extracting features, extracting the feature information of the given target ;Model initialization, use the first frame target features to train the initial model; candidate area, obtain the possible position of the current frame target according to the target position of the previous frame; Feature extraction, perform feature extraction on the target candidate area; target positioning, the model determines that the target is in the candidate The precise location of the area; model update, using online learning to update model parameters to adapt to changes in the appearance of the target; the target tracking method of the present invention provides a model adaptive update method, which can effectively reduce the drastic changes of the model, Prevent tracking frame drift and increase tracking stability.

Description

Technical field [0001] The present invention is an online learning field and a single target tracking field, especially a target tracking method based on online learning. Background technique [0002] The discriminating target tracking method turns tracking problems into a classification problem, and distinguish between targets and backgrounds through training classifiers. In the current frame, the target area is a positive sample, the background area is the negative sample, and the target and background are judged by the online training classifier through the machine learning method. The next frame uses a trained classifier to find the optimal area. The algorithm uses sample image training filter to establish a target appearance model. In the first frame, the primary target window and random imitation transform are performed. A set of sample images are used to train the filter. Then in the subsequent frames The window performs related operations, and finds the maximum value posi...

AI Technical Summary

Problems solved by technology

However, the prediction results generated by the common batch learning method are only based on the once-determined training data set.

[0004] In the process of target tracking, when the target is occluded, the traditional filter still learns the information of the occluder at a fixed model update rate, resulting in the filter being polluted

Images