A pedestrian re-identification method combining deep learning and metric learning

Abstract

The invention discloses a pedestrian re-identification method combining deep learning and metric learning. The pedestrian re-identification method comprises the following steps of extracting pedestrian image deformation area deep convolution features; calculating a pedestrian image deformation area Mahalanobis distance measure and a Jackalide distance measure; and optimizing the distance metric across the camera pedestrian. According to the invention, a region naming network is used; a local deformation area of the pedestrian image is obtained, and the deep convolution characteristic of the local deformation area of the pedestrian appearance is fully applied. The pedestrian recognition method based on the simulated annealing algorithm is used for representing multiple pieces of detail information of the pedestrian image, the simulated annealing algorithm is applied to learn the optimal distance function of the cross-camera pedestrian by combining the Mahalanobis distance and the Jackddistance of the deformation area of the pedestrian image, and the optimal distance measurement of the cross-camera pedestrian is realized, so that the pedestrian re-recognition ability and robustnessare improved.

Description

technical field [0001] The invention relates to the technical field of machine learning and pattern recognition, in particular to a pedestrian re-identification method combining deep learning and metric learning. Background technique [0002] Large-scale intelligent video surveillance systems are widely used in public security, intelligent transportation, national defense and military, and are playing an increasingly important role in improving urban security management and maintaining social stability. In addition to meeting the basic requirements of image / video data collection, transmission, storage and display, the system also needs to have the intelligent analysis function of image / video data. Pedestrian re-identification is one of the important intelligent video analysis technologies. Its task is to let the computer judge whether the pedestrian images appearing in different camera fields of view are the same pedestrian target. [0003] Due to the variability of monitor...

AI Technical Summary

Problems solved by technology

However, most hand-extracted features, such as color, texture, shape features, etc., are either not discriminative enough or not robust to viewpoint changes when performing cross-camera pedestrian matching.

Although deep convolutional features make up for the shortcomings of the above-mentioned manually extracted features to a certain extent, the performance of pedestrian re-identification is seriously affected due to the misalignment of the corresponding position features of pedestrian images under different cameras.

The distance metric learning is from the perspective of optimizing the feature distance metric. Although it can alleviate the appearance difference of different cameras when matching pedestrians to a certain extent, it is difficult to obtain the generalization ability of the Markov model by only using the limited training data on the same data set. distance metric model

In addition, due to the significant changes in the appearance of pedestrians under different cameras, the initial sorting results generated when directly applying the Mahalanobis distance metric obtained during the training process to calculate the characteristic distance of pedestrians may not be accurate enough, thus affecting the performance of pedestrian re-identification

Images