Visual tracking method and device for multiple moving targets

Abstract

The invention relates to a multi-moving-target visual tracking method and device, and the method comprises the two steps: model offline pre-training and online visual tracking. An intra-class loss function with intra-class constraint is fused on the basis that an original loss function has inter-class constraint to form a total loss function, and the total loss function replaces the original loss function to carry out offline training on a model. Therefore, the inter-class feature vector spacing between different classes of targets and the intra-class feature vector spacing between different entities of the same class are increased at the same time. Through the technical scheme disclosed by the invention, on the basis that a visual tracking task does not increase the training burden of a network model and the network model has relatively good feasibility and real-time performance, the defect that the distinguishing of different entities of the same category is insensitive in the prior art is overcome, and the visual tracking accuracy of multiple moving targets is improved.

Description

technical field [0001] The invention belongs to the technical field of computer vision tracking, in particular to a multi-moving target vision tracking method and device. Background technique [0002] Modern computer vision tracking usually performs network model pre-training under the image classification framework. This training method encourages the extraction of inter-class features between different categories of objects, and is insensitive to the intra-class features between different target entities of the same category. However, One of the problems of visual tracking is the interference of similar targets. That is, in a scene with a large number of similar instances of the same type, the interference of similar targets often causes the tracker to drift, which brings great difficulties to visual tracking. A method that does not increase the network training burden, A visual tracking method that improves intra-class feature extraction for different target entities of t...

AI Technical Summary

Problems solved by technology

[0003] The purpose of the present invention is to provide a multi-moving target visual tracking method and device. By optimizing the loss function for network training, the network model can simultaneously increase the feature distance of different types of targets and different target entities of the same type, so as to solve the problem under the current image classification framework. The defect that the feature discrimination of different target entities of the same type is not high in multi-target visual tracking

Images