RGB-D feature target tracking method based on twin network

Abstract

The invention discloses an RGB-D feature target tracking method based on a twin network. The method comprises the following steps: constructing a twin network model based on RGB-D features; precessing the template image by a shared network to obtain semantic features of the template image, and inputting the high-level semantic features to a deep convolutional network module to obtain a depth map; performing depth feature extraction on the depth map to obtain depth feature information, and fusing the depth feature information with the semantic features in a cascade mode to obtain fused image features; processing the search image through a shared network to obtain features of the search image, wherein the features of the search image are subjected to convolution and pooling operation to obtain context information of the search image, the fused image features are guided through the context information of the search image, and adaptive features used for target positioning are generated; and carrying out cross-correlation operation on the adaptive features and features obtained by processing the search images through a shared network, and carrying out interpolation calculation on the score graph to obtain a tracking result. The depth map is introduced, high-precision tracking in a complex scene can be achieved, and the effect is good.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to a twin network-based RGB-D feature target tracking method. Background technique [0002] Target tracking is one of the topics with important research significance in the field of computer vision. It refers to the detection, extraction, identification and tracking of moving targets in image sequences, and obtains the parameters of moving targets for further processing and analysis. Achieving behavioral understanding of moving objects for more advanced detection tasks. The target tracking method based on deep learning can be summarized into three types according to the network function: deep target tracking method based on correlation filter, deep target tracking method based on classification network and deep target tracking method based on regression network. [0003] Depth target tracking method based on correlation filtering: This method trains the correlation filter f...

AI Technical Summary

Problems solved by technology

But this method needs to rely on large-scale training data to ensure the robustness of the model

Images