Person reidentification (Re-ID) is a computer vision problem that refers to the process of identifying a person from multiple camera views or across different scenes. Person ReID enables cross-camera identification by matching features of the same individual from nonoverlapping camera views or different times, addressing limitations of traditional surveillance with applications in security, retail analytics, smart cities, and healthcare. Despite its advancements and wide use, Re-ID faces challenges that include variations in appearance, lighting and camera angle differences, and occlusion in crowded areas. This paper proposes and develops a novel deep learning approach called “SiamID” that leverages multiperspective views of individuals and intergates keypoints rather than relying solely on body parts or facial features for person reidentification. Existing person Re-ID datasets, such as CUHK03 and DukeMTMC-ReID, have offered diverse and challenging scenarios. However, they often overlook issues like image bifurcation based on varying orientations and keypoint extraction. To address these challenges, the proposed work introduces a new dataset called ‘SNK-PU’ that improves Re-ID research in people through better classification of orientation and keypoint analysis. This data set includes images of 74 subjects, each captured from four different angles: front, back, left, and right. These varied viewpoints improve the ability to identify people in different perspectives on the camera. In addition, this data set incorporates the height of each person, which further assists in accurate identification. The proposed approach demonstrates good accuracy on existing datasets, achieving 94.53% on Cuhk03, 90.42% on Duke, 94.53% on Market1501, and 97.65% on Mars while taking complete dataset for training. For our novel SNK-PU dataset, the proposed approach achieves 86.45% accuracy for partial categories and 95.11% for complete categories using a Siamese Network. These results highlight the robustness of our method, particularly in scenarios where precise person re-identification is critical. Additionally, the uniqueness of our approach effectively addresses challenges related to occlusion and viewpoint variance in person reidentification.
Sohanvi et al. (Fri,) studied this question.