Target tracking method for RGB-D (RGB-Depth) data cross-modal feature learning based on sparse deep denoising autoencoder

Abstract

The invention discloses a target tracking method for RGB-D (RGB-Depth) data cross-modal feature learning based on a sparse deep denoising autoencoder. The method comprises the steps of establishing a RGB-D cross-modal feature deep learning network with a sparse limitation denoising autoencoder; training the RGB-D cross-mode feature deep learning network by employing an unsupervised learning method; inputting RGB image positive and negative samples and Depth image positive and negative samples into the trained RGB-D cross-modal feature deep learning network, thereby obtaining cross-modal features; sending the cross-modal features to a logic regression classifier, training the logic regression classifier by employing a supervised learning algorithm; generating a particle set according to a state transition model; sending the cross-modal features of each particle to the trained logic regression classifier, thereby obtaining confidence scores as observation likelihood modals; and obtaining a posteriori probability of a tth frame in a RGB-D video sequence by employing a particle filtering method, thereby obtaining target tracking results X<^><t>. The method is relatively high in accuracy.

Description

technical field [0001] The invention relates to a target tracking method, in particular to a target tracking method based on RGB-D data cross-mode feature learning of sparse depth denoising self-encoder. Background technique [0002] Video object tracking technology is one of the key technologies in the field of computer vision, and it has been widely used in intelligent video surveillance, robot navigation, video measurement and other fields. The video target tracking methods in the prior art mainly realize target positioning in a two-dimensional image sequence. Due to the lack of three-dimensional information, tracking failures will occur when the target is occluded, rotated, or changes in attitude. With the introduction of depth sensors such as Microsoft Kinect, Asus Xtion, and PrimeSense, it has become a reality to obtain color images and depth images (RGB-D, RGB-Depth data) at the same time. Depth information can improve the performance of video object tracking. How to...

AI Technical Summary

Problems solved by technology

Depth information can improve the performance of video object tracking. How to better integrate RGB information and depth information to improve the performance of video object tracking is the focus of research in the field of video object tracking. The following technical problems generally exist in previous related technologies: 1. Both the HOG feature and the Harr feature are artificially designed features, which have certain limitations; 2. These related technologies operate on the artificially designed features in RGB mode and Depth mode (depth mode), and then use weighted and other methods simply fuse the features extracted in the two modes, and the fusion method ignores the complex correlation between the RGB mode and the Depth mode

Images