Enhanced multilayer convolutional visual tracking method

Abstract

The invention discloses an enhanced multilayer convolutional visual tracking method. The method comprises the specific steps that a VGG-19 deep network framework is adopted, a 0.5-time third convolutional layer and a 0.5-time fourth convolutional layer are directly added to serve as a feature extraction layer, a feature template is output, and dimension reduction is performed on the feature template through principal component analysis to obtain a needed feature map; added convolutional features are adopted, and the robustness of extracted features is enhanced; according to a target position,the ratio and weight of a color histogram are determined, the weight is multiplied with target size of the last frame, and the product is current target size; whether a target vanishes is judged according to the judgment of whether a maximum response value output by a filter template is greater than a given threshold; if the maximum response value is greater than the given threshold, the target position is directly determined; and if the maximum response value is smaller than the given threshold, the target vanishes, a to-be-detected target position in a detection region is determined throughan interval frame difference method, non-target object interference is eliminated through the histogram weight, a suspicious target object is determined, features are extracted and correlated with thefilter template, a maximum response greater than the threshold is found out, and the target position is determined. Through the method, the problem that poses, illumination intensity and other factors cause target appearance changing, sheltering, etc. during tracking is relieved.

Description

technical field [0001] The invention relates to the field of machine vision target tracking, in particular to an enhanced multi-layer convolution vision tracking method. Background technique [0002] Machine vision has attracted more and more attention in recent years. As an important research direction of machine vision, target tracking has a wide range of applications in intelligent robots, intelligent transportation, and human-computer interaction. At present, target tracking algorithms are mainly divided into two types: generative tracking method and discriminative tracking method; generative target tracking algorithm focuses on the description of the target itself, and continuously searches for the most similar area to the target. Representative methods include template matching, Particle filter, mean shift algorithm, etc. The discriminative method aims to distinguish the target from the background, which is to change the tracking problem into a binary classification p...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to propose an enhanced multi-layer convolution visual tracking method, which extracts more robust convolution features through a multi-layer convolution network under the framework of a fast adaptive correlation filtering algorithm, adds a scale strategy, and introduces frame difference method and edge Extraction to solve the problem of occlusion and rapid target changes in tracking

Images