A fine-grained semantic color person re-identification method based on the position constraints of human body parts

Abstract

The invention discloses a fine-grained semantic color pedestrian re-identification method based on human body part position constraints, which belongs to the technical field of surveillance video retrieval. The invention implements semantic color-based pedestrian re-identification by introducing fine-grained color representation and human body part position constraint relations. Improvement of recognition effect. Firstly, the bag-of-words model (BOW) in image retrieval is combined with the semantic color feature representation method, so as to refine the color range and increase the color types. The present invention calls it a fine-grained color model. Secondly, three kinds of refined human body position constraints are introduced into the fine-grained color representation model, which are position weight, upper and lower constraints, and drift correction, which are mainly implemented in three forms: Gaussian template, divided into horizontal stripes, and overlapping levels formed by sliding windows. stripe. The effectiveness of each step is effectively demonstrated on the VIPeR and CUHK datasets. At the same time, it shows that semantic features are a good supplement to visual features, and can further improve the effect of pedestrian re-identification.

Description

technical field [0001] The invention belongs to the technical field of surveillance video retrieval, and in particular relates to a fine-grained semantic color pedestrian re-identification method based on position constraints of human body parts. Background technique [0002] Pedestrian re-identification refers to the technology of judging whether pedestrian images appearing under different surveillance cameras belong to the same pedestrian. Because pedestrian images under multiple cameras often have changes in perspective, illumination, posture, and size, the differences between the same pedestrians are even greater than those between different pedestrians. Existing person re-identification techniques can be roughly divided into two categories: person re-identification techniques based on feature representation, and person re-identification techniques based on scale learning. However, the person re-identification technology based on scale learning relies on a large number ...

AI Technical Summary

Problems solved by technology

However, existing color-based feature description methods only use limited color categories (11 or 16) to describe pedestrians (called coarse-grained color names, such as [Document 3], [Document 4], [Document 5]) , some discriminative information is thus lost

Images