A video pedestrian re-identification method based on multi-attention heterogeneous network

Abstract

The invention discloses a video pedestrian re-identification method based on a multi-attention heterogeneous network, belonging to the field of image processing. The method includes: constructing and training a multi-attention heterogeneous network; using the trained network to extract features from videos with known pedestrian IDs and videos with undetermined pedestrian IDs, and determine pedestrian IDs according to the cosine distance between the two features. The present invention introduces Soft attention and non-local attention into the OSNet network, uses Soft attention to focus on pedestrian area features in images, uses non-local attention to learn the spatio-temporal features of video sequences, and improves the feature representation of video sequences , to extract more robust and discriminative features and improve the accuracy of recognition. At the same time, the features of a specific frame are selected as the branch of the local feature learning network. While learning the global features of pedestrians in the video sequence, the learning of the local features of pedestrians is strengthened, and the performance of the network in video pedestrian re-identification is improved.

Description

technical field [0001] The invention belongs to the research field of pedestrian re-identification in image processing and machine vision, and more particularly, relates to a video pedestrian re-identification method based on a multi-attention heterogeneous network. Background technique [0002] Pedestrian re-identification is a basic task in automatic video surveillance and a research hotspot in recent years. Video-based person re-identification aims to match video sequences of pedestrians on non-overlapping cameras. In order to realize video-based person re-identification, the typical method is to learn a mapping function to project the video sequence into a low-dimensional feature space, and then determine the pedestrian ID by comparing the distance between samples. [0003] Numerous studies have demonstrated that convolutional neural networks have surpassed traditional handcrafted features as a mapping function, and then aggregated image features by average pooling or m...

AI Technical Summary

Problems solved by technology

Although these methods successfully capture the spatial and temporal features of video sequences, they only explore high-level feature aggregation for pedestrian feature representation, which may not be robust enough for fine-grained classification tasks such as video person re-identification.

[0004] In recent years, more and more attention models have been applied to various fields of computer vision. The attention model can focus on learning the most informative part of the input signal, and can effectively improve the network's ability to learn pedestrian features. However, the existing Algorithms often only use a single attention, ignoring the heterogeneous characteristics and complementary effects of different attention networks, so that the discriminativeness of the extracted pedestrian sequence features is relatively low, and the accuracy of video pedestrian re-identification is low

Images