Target tracking method based on attention cycle network

Abstract

The invention discloses a target tracking method based on an attention cycle network. A local position attention mechanism and an appearance attention mechanism are introduced into a target tracking framework; three depth models are set up, a cyclic convolutional neural network is adopted to perform time sequence prediction; the uncertainty evaluation mechanism and other technical means are addedinto the tracking framework, the efficiency and accuracy of visual target tracking based on the computer algorithm are greatly improved, high reliability and promotable value are achieved, and compared with other time sequence prediction methods, the parameter amount is small, the speed is high, and the accuracy is high; an uncertainty evaluation mechanism is used in the tracking process, the quality of a tracking result can be guaranteed, a tracker is initialized or tracking is stopped in time when the quality is reduced, too many wrong results are prevented from being given, and higher reliability is achieved.

Description

technical field [0001] The invention relates to the technical field of visual target tracking algorithms, in particular to a target tracking method based on an attention loop network. Background technique [0002] Target tracking is one of the important problems in computer vision. The main purpose is to track multiple targets in the video screen and give the target's trajectory; the typical scene of target tracking is: for continuous video sequences, artificially given one or more Target, find and distinguish multiple calibrated targets in subsequent video frames; [0003] The algorithm model of computer vision for target tracking is mainly divided into two types: generative model and discriminative model, among which: [0004] ①Generation model: learn the joint probability distribution of data, judge by finding the conditional probability distribution, and learn the way of data generation; [0005] ② Discriminant model: directly learn the conditional probability distribu...

AI Technical Summary

Problems solved by technology

[0009] Disadvantage 1: The timing feature of the target is an important feature in target tracking. At present, most methods only consider the appearance feature of the target, and assume that the appearance feature of the tracking target does not change with time, which loses important priors in the target tracking problem;

[0011] Disadvantage 2: Some target tracking methods that introduce time information use long-short-term memory neural network (LSTM) for timing prediction, which has a large amount of parameters, slow speed and cannot predict image information;

[0013] Disadvantage 3: Most target tracking methods do not have a tracking result evaluation mechanism, and cannot give the model's confidence in the prediction results, so it is difficult to apply in occasions that require high reliability;

Images