Occluded pedestrian re-identification method combining spatial transformation network and multi-scale feature extraction

Abstract

The invention discloses an occluded pedestrian re-identification method combining a spatial transformation network and multi-scale feature extraction. The method comprises the following steps: constructing an occluded pedestrian picture set by using a simulated occlusion generator; forming a data set by the original picture and the shielded pedestrian picture, and inputting the data set into a spatial transformation network for spatial transformation correction; performing multi-scale feature extraction on the corrected images through a convolutional neural network and a spatial pyramid pooling layer, and combining the images into a fixed-length one-dimensional feature vector; obtaining a one-dimensional feature vector containing K elements from the fixed-length one-dimensional feature vector through a full connection layer, and performing identity classification training to obtain a trained network; and extracting features of a pedestrian image to be queried by using the trained network and carrying out similarity matching. According to the method, multi-scale feature extraction is carried out, and the model is more robust by combining feature maps of different scales; and a spatial transformation network is introduced and can be directly embedded into any depth network model for end-to-end training.

Description

technical field [0001] The invention belongs to the technical field of deep learning and computer vision, and in particular relates to a re-identification method for occluded pedestrians combined with spatial transformation network and multi-scale feature extraction. Background technique [0002] Pedestrian re-identification is considered to be a sub-problem of image retrieval. It hopes to use computer vision technology to track across cameras. Specifically, given the pedestrian image to be queried under a certain camera, the image library obtained from other non-overlapping cameras Pedestrian images with the same identity are retrieved from . This technology is widely used in public safety fields such as video surveillance and intelligent security. In the past few years, the problem of pedestrian re-identification has been widely studied, but they usually assume that the image database and the image to be queried are complete pedestrian images, and our pedestrians will ine...

AI Technical Summary

Problems solved by technology

This method can greatly improve the accuracy of recognition, but due to the simultaneous graph network and human key point estimation, the model is more complex, and the training cost is higher

Images