An integrated method for person re-identification based on metaclass base learner

Abstract

The invention discloses a pedestrian re-identification integration method based on a metaclass base learner. The method of the present invention firstly clusters the depth feature vectors of pedestrian images to obtain a set of metaclasses, which can make the images in the metaclasses have similar semantics in the feature space; obtain the information corresponding to each metaclass by evenly dividing the fully connected layer of the convolutional neural network. The base learner solves the pedestrian re-identification sub-problem corresponding to the metaclass; at the same time, the sampling process of the positive and negative example pairs within the metaclass and the negative example pairs between metaclasses, the training scheme of the base learner and the adaptive Gradient-weighted base learner differentiation strategy. The present invention can effectively deal with the complexity of pedestrian image data distribution, efficiently utilize the representation space of depth features, and obtain better pedestrian recognition performance.

Description

technical field [0001] The invention belongs to the field of computer technology, in particular to the technical field of pedestrian re-identification in computer vision, and relates to a pedestrian re-identification integration method based on a metaclass base learner. Background technique [0002] With the construction of digital cities, video surveillance equipment is widely used in public places such as roads, commercial areas, and living areas. How to intelligently analyze the collected massive video data has become particularly important. Person Re-identification (Person Re-identification) is an important applied research problem whose goal is to match images of persons of the same identity across different camera views. It can be widely used in intelligent video surveillance, security, criminal investigation and other fields, such as pedestrian tracking across camera perspectives, pedestrian search in large image galleries, visitor analysis in retail stores and other ...

AI Technical Summary

Problems solved by technology

[0004] To sum up, the deficiencies of existing person re-identification methods are mainly manifested in the following aspects: First, all dimensions of the depth feature vector are directly used as a single metric learner, resulting in low efficiency of feature representation space and pedestrian The recognition performance is not high; second, the data distribution of the entire dataset is directly fitted, and the complexity of the pedestrian image data distribution is not fully considered, which usually leads to local optimum and overfitting in practice; third, due to the use of different metrics There are differences in the base learners of the indicators, which usually cause the problem of slow model convergence

Images