Deep integration target tracking method based on time and space network

Abstract

The invention discloses a deep integration target tracking method based on a time and space network. The method comprises: using a VGG-16 network to extract characteristics of third, fourth, and fifthlayers, expressing the traditional correlation filtering algorithm as a layer of convolutional neural network to obtain a correlation filtering network; a time network and a spatial network are constructed on the basis of a related filtering network, so that time information and space information of the target are further captured, improving the accuracy of an algorithm. Meanwhile, all weak trackers are fused into strong trackers through an adaptive weight integrated learning algorithm to achieve target tracking, integrated learning enables the method to have good robustness, target trackingunder complex scenes can be coped with, finally, an updating strategy combining short-time updating and long-time updating is provided, and the stability of the model is ensured.

Description

technical field [0001] The present invention relates to the technical fields of image processing and computer vision, and specifically relates to a deeply integrated target tracking method based on time and space networks. Accurate tracking of targets is achieved through feature extraction, construction of deep networks, and integration of weak trackers. It is used in vehicle navigation, human-computer interaction and video surveillance and other fields. Background technique [0002] Visual object tracking is a basic problem in the field of computer vision and can be widely used in many practical systems such as vehicle navigation, video surveillance, and human-computer interaction. At the heart of the problem is how to develop a robust appearance model with extremely limited training data (usually bounding boxes in the first frame). In the past few decades, visual target tracking technology has made great progress, mainly including tracking methods based on correlation fil...

AI Technical Summary

Problems solved by technology

However, the above correlation filtering algorithm has two limitations.

First, learning correlation filters and feature extraction are independent of each other, that is, there is no end-to-end training model

Second, the update method of most correlation filtering algorithms is relatively simple. Basically, linear interpolation is used to update the learned filter to achieve the effect of model adaptation. In fact, this method is only an empirical operation. Once there is noise to update, cause the target to drift

But for the target tracking task, it is not enough to use the features of the last layer to represent the target, because it also needs to accurately locate the target

Images