Single-target tracking method based on multiple networks

Abstract

The invention provides a single-target tracking method based on multiple networks, and the method is a model adopting deep learning. The method comprises the following steps: cutting a first frame image and a current frame image of a video sequence to obtain a template image and a to-be-searched image; inputting the template image and the to-be-searched image into an appearance subnet and a semantic subnet, respectively obtaining low-level appearance features and high-level semantic features of the template image and the to-be-searched image, and performing feature fusion to respectively obtain fusion feature maps of the template image and the to-be-searched image; then, based on the fusion feature map of the template image and the image to be searched, using a similarity discrimination method to obtain a final response map; and finally, obtaining a tracking result according to the information provided by the final response diagram. According to the method, the problems that a traditional single-target tracking method cannot effectively detect a tracking target in a to-be-searched image containing a similarity background and false detection is caused by noise existing in extractedlow-layer appearance features in a feature extraction method based on deep learning are solved.

Description

technical field [0001] The invention belongs to the field of image processing and computer vision, in particular to a single target tracking method based on multiple networks. Background technique [0002] Single object tracking is an important part of the information fusion method, which is widely used in many fields such as video surveillance, virtual reality, human-computer interaction, and unmanned driving. The essence of single target tracking is to estimate the continuous motion state of a single dynamic target through filters. The general framework of single target tracking is to first initialize the features of the target to be tracked by extracting the features of the first frame input, construct the target model, then extract the features of the current frame, and perform similarity with the features of the tracking target in the first frame Judgment, and finally output the position estimate of the tracking target in the current frame. [0003] Single target trac...

AI Technical Summary

Problems solved by technology

[0006] The problem to be solved by the present invention is: the features used in the traditional single target tracking method are not accurate enough to form robust features to the target, and the model with better tracking effect is slower

Images