关键词: 全局, 局部, 跨维度多尺度池化融合, 多层级注意力, 高低频特征

Abstract: Person re-identification (ReID) is a technology that utilizes artificial intelligence to solve public safety application issues such as station security checks and urban surveillance systems, with the capability to identify specific individuals from images collected across different devices. However, in problems like person re-identification, factors such as deliberate occlusion of pedestrians or occlusions caused by complex scenarios significantly increase the difficulty of person re-identification. In most existing occluded person re-identification methods, convolutional neural network (CNN) models tend to focus more on local features but struggle to obtain global structural information, while Transformer network models excel at modeling long-range feature dependencies but often overlook the details of local features. To address these challenges, this paper proposes an occluded person re-identification method based on global-local fusion features. This method leverages the characteristics of CNN and Transformer feature learning networks to enhance the global representation ability of features while enriching pedestrian local features. The proposed model consists of three parts: the CNN network primarily extracts local detail features, while the Transformer branch focuses on extracting global feature information. A cross-dimensional multi-scale pooling fusion module calculates the correlation between the features of the two branches, thereby achieving global-local feature fusion. A mask generation module guided by multi-level attention can accurately highlight key features in pedestrian images, automatically align pedestrian feature information, and suppress interference from occluded parts or background noise. Image high-frequency and low-frequency feature enhancement module strengthens the high and low frequency feature information of occluded pedestrians, highlighting effective information. The ablation experiments and experimental results on relevant datasets prove the effectiveness of the proposed method.

Key words: global, local, cross-dimensional multi-scale pooling fusion, multi-level attention, high and low frequency features