Multi-Model Person Re-identification: Combining Facial and Body Features

Abstract

In the surveillance industry, Person Re-identification (Re-ID) is significant as it matches a person�s appearance across multiple non-overlapping cameras. How-ever, this task is challenging due to changes in camera viewpoints, occlusion, and varying appearances such as clothes, shoes, and pose. To overcome these challenges, discriminative feature learning is necessary. Recently, For this aim, deep convolutional neural networks (CNNs) have been widely employed. This study proposes a lightweight and robust network for person Re-ID, which employs the YOLOv4 object detection model for pedestrian detection and the DeepSORT algorithm for tracking. The proposed model is designed to learn discriminative features at multiple semantic levels, utilizing the ResNeXt architecture as a back-bone. Specifically, the network comprises multiple blocks, where channels are concatenated between blocks, and an aggregation gate is used to aggregate the output of multiple channels. The aggregation gate produces channel-wise weights that dynamically fuse the resulting multi-scale feature maps. This layout effectively allows the model to extract discriminative features even under challenging condi-tions. To determine whether our recommended strategy is effective, we conducted experiments on the widely-used Market1501 dataset and our own custom-made datasets, including indoor, outdoor, and same-dress outdoor datasets, which cover various challenges such as occlusion, lighting variations, and similar body fea-tures. The experiment results show that our approach to Person Re-identification is effective in Person Re-identification. These results indicate the potential of our proposed method to be applied in various real-world scenarios, such as surveil-lance cameras in markets, shopping malls, parking areas, and other public places.