SNK-PU: Leveraging Multi-View CCTV Footage for Improved Person Re-Identification

Person re identification (ReID) is a crucial computer vision task aimed at matching individuals across different camera views despite variations in viewpoint, lighting, and occlusions. Even though CCTV cameras are everywhere, person reidentification remains challenging due to their inefficiency in accurately capturing faces when they are covered or blurred, as most cameras are optimized for capturing faces from a single direction. The lack of datasets that focus on multi-view capturing of individuals from all four sides makes this issue worse, with few models addressing the need to reidentify individuals based on multiple images from various angles. To tackle this problem, we have employed a novel technique that leverages the keypoints and gait of individuals from a multi-dimensional perspective, rather than relying solely on facial or biometric images. Our proposed novel SNK-PU dataset addresses this gap by including images of individuals from the four specific orientations: front, back, left, and right. Additionally, it incorporates supplementary parameters such as height and age for ReID. We trained a Siamese model with a triplet loss function on the proposed SNK-PU dataset, utilizing our innovative image generation technique. When the entire dataset is used for both training and testing, the model achieves an accuracy of 95.11%. However, with an 80:20 train-test split, the accuracy reaches 86.45%, reflecting a more realistic evaluation scenario. Unlike traditional methods that rely on uniform image recognition, our approach leverages easily accessible keypoints, enabling more consistent and robust person re-identification across diverse conditions.