An object tracking method based on deep learning features and point-to-set distance metric learning

Abstract

The invention discloses a target tracking method based on deep learning features and point-to-set distance metric learning, comprising the following steps: randomly selecting a number of target samples and background samples in the initial frame of tracking; performing target sample feature extraction on the target samples, Extract background sample features from background samples; cluster the extracted target sample features into several target template sets, and cluster the extracted background sample features into several background template sets; reduce the distance between samples of the same category and increase the difference The distance between samples is used to learn the projection matrix; the target candidate is collected for subsequent frames according to the Gaussian distribution; the features of the target candidate are extracted, and the target template set, the background template set and the target candidate are projected into a common subspace using the projection matrix; The distance from each target candidate to all target template sets, the sum of the distances is used as the score of each target candidate, and the final tracking result is the average value of the first few target candidates with the smallest distance.

Description

technical field [0001] The invention relates to the technical fields of image processing and pattern recognition, in particular to a target tracking method based on deep learning features and point-to-set distance metric learning. Background technique [0002] Object tracking is an important research direction in the field of computer vision, and it has a wide range of applications in video surveillance, virtual reality, human-computer interaction, automatic driving and other fields. Currently, discriminative tracking methods have achieved better tracking results. Most discriminative tracking methods regard target tracking as a classification task, select target samples and background samples in the first frame to train an SVM classifier; for subsequent frames, collect several target candidates in each frame, and each target candidate is The classifier is classified as object or background; the candidate with the largest object confidence is recorded as the tracking result....

AI Technical Summary

Problems solved by technology

Since the SVM classifier is only based on a small number of support vectors (that is, a small number of samples selected from the training samples as the classification boundary) and the samples are linearly inseparable in most cases, this ignores the role of the remaining samples in the classification process

Images