Pedestrian Attributes Recognition for UAV-Human

Pedestrian Attributes Recognition (PAR) aims to predict a set of attributes from a predefined list to describe the characteristics of a pedestrian. This paper addresses the challenge of attribute recognition in the UAV-Human dataset. In this dataset, there are two key factors constraining attribute recognition performance. Firstly, there is class imbalance in each attribute, and the distribution of classes differs significantly between the testing and training sets. Secondly, some attributes cannot be observed from the current perspective of the images. For example, we cannot observe the attribute of the backpack from a frontal image. To tackle these issues, this paper proposes Class Imbalance Correction Strategy (CIC) and ID-based Attribute Correction Strategy (IAC) to enhance attribute recognition performance. CIC balances training classes using undersampling and oversampling to match the class distribution in the testing set. IAC corrects misidentified attributes by leveraging person re-identification (ReID) techniques to obtain images of the same pedestrian from different viewpoints. Our approach significantly improves the accuracy of attribute recognition over the baseline method. The code is publicly available at https://github.com/laiping-lp/MMVRAC-Attribute-Recognition.

I. Introduction

Pedestrian attribute recognition (PAR) has attracted more and more attention due to its wide application scenarios. Given a single image of a person, the goal of the PAR task is to recognize a set of semantic attributes (for example, age, gender, and clothing) that describe the characteristics of that person [1]. PAR has great application potential in the field of intelligent video surveillance and security. Pedestrian detection [2], pedestrian retrieval [3], and person re-identification [4]–[7] can also be used as auxiliary means. PAR is a challenging task because the model is significantly affected by viewpoint, light, and resolution [8].