Adaptive compressive tracking algorithm based on multiple features

Abstract

The invention belongs to the field of visual tracking and discloses an adaptive compressive tracking algorithm based on multiple features. Firstly, a tth frame of image is extracted and is grayed, a compressive sensing theory is used, extracted high-dimensional and multi-scale image features are mapped to low-dimensional subspace respectively through two extremely-sparse random matrixes, two low-dimensional image features are obtained, a classifier is used for classifying the features, and the sample with the maximum reaction value of the classifier is selected to be the tracked target. Then, the Hamming distance between a Hash fingerprint of the current frame of tracked target and a Hash fingerprint of the previous frame of tracked target is calculated, the Hamming distance is compared with a high-low threshold, and the updating degree parameters of the classifier are adjusted adaptively. The algorithm has high robustness and high accuracy under various interference factors, the average speed for image processing can achieve 19.8 frames per second, real-time performance requirements can be achieved basically, and an important practical significance is realized.

Description

technical field [0001] The invention belongs to the field of visual tracking, and relates to an adaptive compression tracking algorithm based on multiple features. Background technique [0002] Object tracking is a very popular research topic in the field of computer vision because of its significance in vehicle navigation, traffic monitoring, and human-computer interaction. Although the subject of object tracking has been studied for decades and many tracking algorithms have been proposed, it is still a very challenging problem. Because the target appearance is disturbed by various factors, such as illumination changes, pose changes, complete or partial occlusions, and sudden movements, etc. Therefore, it is a challenging problem to develop a high-performance tracking system under the interference of the aforementioned factors. [0003] CT (Real-time Compressive Tracking) tracking algorithm is a simple and efficient real-time compression tracking algorithm proposed by K. ...

AI Technical Summary

Problems solved by technology

However, since the feature collected by the tracking algorithm is single, it cannot adapt to the rapid movement of the target and environmental changes, which in turn causes problems such as target drift and tracking instability.

In addition, an update degree parameter is introduced in the update process of the classifier in the tracking algorithm, but this parameter is given in advance and is fixed in the whole process of target tracking. If the given value is too large, the classification The update process of the classifier cannot adapt to the change of the target appearance, and the given value is too small to make the classifier "over-learning"

Images