Pedestrian re-identification method based on clustering guidance and paired measurement triple loss

Abstract

The invention discloses a pedestrian re-identification method based on cluster guidance and paired measurement triple loss. The method comprises the following steps: 1, acquiring a similarity matrix based on features output by a deep learning network and corresponding labels; 2, calculating sampling loss of hard cosine similarity measurementsamples in paired measurement; 3, calculating sampling loss of hard European style similarity measurement samples in the paired measurement; 4, calculating a clustering guidance correction term, and fusing all the losses to obtain clustering guidance and paired measurement triple losses; and 5, combining clustering guidance and paired measurement triple loss with cross entropy loss based on representation learning to obtain final loss, and adding the final loss into network parameter training for updating. According to the method, in combination with a paired measurement mode, a deep learning model can complementarily mine the similarity of the samples from different angles; the similarity between the samples is maximized through a correction item guided by clustering; and finally, the method is applied to deep learning training of pedestrian re-recognition to improve the performance of the model.

Description

technical field [0001] The invention belongs to the field of computer vision and pattern recognition, and relates to a pedestrian re-identification method based on similarity clustering guidance and pairwise similarity measurement triplet loss. Background technique [0002] In recent years, due to the widespread use of large-scale multi-camera surveillance systems in public places (campus, shopping malls, airports, hospitals, etc.) and the demand for intelligent surveillance and security systems, technologies such as pedestrian detection and pedestrian re-identification have developed rapidly. In the field of computer vision, person re-identification (Person Re-identification) is considered to be the next advanced task in pedestrian tracking system, which aims to predict the identity correspondence of different multi-camera surveillance. At present, the method of deep learning is mainly used to solve the problem of pedestrian re-identification, and the representation learnin...

AI Technical Summary

Problems solved by technology

[0004] 1. The earlier work improved the triplet loss function by exploring the correlation between the features extracted by the deep network, but these works were not compatible with the existing neural network due to the long time;

[0005] 2. The newer work achieves self-regulation and self-learning by introducing additional weight factors in the triplet loss function, but the newly introduced parameters will increase the computational complexity of the model and the difficulty of parameter adjustment for network training settings

The present invention introduces a paired measurement method in the feature space, and complementarily quantifies the similarity of features in the Euclidean distance and cosine distance respectively, but the paired measurement mainly improves the accuracy between different samples of the same label, and cannot solve abnormalities The problem with too many values

Images