Efficient object tracking method based on multi-branch autoencoder adversarial network

Abstract

Efficient object tracking method based on multi-branch autoencoder adversarial network. Collect a large number of target templates and search area sample pairs containing targets in the labeled offline target tracking dataset; use the mean square error loss to conduct preliminary training on the proposed target probability generator in a fully supervised manner; introduce a discriminator, Add the method of adversarial training to jointly optimize the target probability generator and discriminator; given the first frame in the test video, sample the marked target area as the initial target template; given the test frame, use N times the size of the current target length and width The window is randomly displaced to obtain the search area; the search area and the target template are input into the target probability generator, and the target probability map is output, and the position of the maximum point in the target probability map is selected as the target center; the target is estimated according to the distribution of the target probability map At the scale of the current frame; update the target template according to the estimated target area of ​​the current frame.

Description

technical field [0001] The invention relates to computer vision technology, in particular to an efficient target tracking method based on a multi-branch self-encoding confrontation network. Background technique [0002] Object tracking is a basic research in the field of computer vision. In many current fields with high real-time requirements, such as vehicle automatic driving and UAV navigation, object tracking plays an extremely important role. Therefore, how to design a robust real-time object tracking method for different tasks in practice is of great significance. [0003] In recent years, deep convolutional neural networks have achieved great success in various applications in the field of computer vision (such as object detection, instance segmentation, etc.). Much of this can be attributed to the availability of a large number of labeled datasets. Deep convolutional neural networks can learn better feature representations from labeled datasets, resulting in far bet...

AI Technical Summary

Problems solved by technology

However, in the field of target tracking, the application of generative adversarial networks is still relatively limited. The main reason is that in target tracking tasks, tracking algorithms can only obtain relatively effective online samples of targets, and limited online samples greatly limit the generation The online learning of the adversarial network prevents it from fully converging

At the same time, online learning will greatly affect the tracking speed of the tracking algorithm

Images