Person Re-Identification in Video Surveillance Across Multiple Cameras Achieving Cross-Domain Generalization

Abstract

Person re-identification (Re-ID) is a critical com ponent of intelligent video surveillance systems, aims to match individuals across non-overlapping camera views. Despite sig nificant progress, performance degradation caused by domain shifts such as variations in illumination, viewpoint, background, and camera characteristics remains a major challenge in real world deployments. This study presents a robust cross-domain person re-identification framework based on the Omni-Scale Network (OSNet) enhanced with Batch Normalization Neck (BNNeck) and MixStyle for improved domain generalization. OSNet effectively captures discriminative multi-scale visual cues, whereas BNNeck decouples feature learning for classification and metric learning, improving embedding discrimination. The proposed model was trained on a source-domain dataset and evaluated on a target-domain dataset without any target-domain supervision. These results confirm the effectiveness of combining omni-scale feature extraction with style-based augmentation for robust person re-identification in multi-camera- video surveillance environments. The key areas where the person re-identification applied in Intelligent video surveillance systems, law enforcement and criminal investigations, smart city monitoring, border control and immigration security, retail analytics and customer behavior analysis, campus and enterprise security, smart transportation hubs, healthcare and assisted living facilities, large-scale event and crowd management, and human–robot interaction in autonomous systems.