MODERN APPROACHES TO ONLINE MULTI-OBJECT SURVEYS BASED ON DEEP METRIC LEARNING AND RE-IDENTIFICATION

This article analyzes modern approaches in the field of Multiple Object Tracking (MOT) based
on deep metric learning and re-identification (Re-ID). The main focus is on solving the problem of heavy occlusion, unsupervised Re-ID, Siamese neural networks and metric learning methods. The analysis was conducted based on five high-impact scientific works, including approaches such as the unsupervised Re-ID module, the Siamese Deep Metric Tracker (SDMT), metric learning based on the von Mises-Fisher (vMF) distribution, and SiamMOT. The results show that appearance-based embeddings (feature vectors based on the visual appearance of the object) and occlusion estimation modules help correctly store object identifiers even under conditions of heavy occlusion, significantly reducing identifier switching (ID switch). These approaches make it possible to accurately assess the similarity of objects and re-identify objects lost under a temporary barrier. However, maintaining a balance between high accuracy and real-time performance, as well as saving computing resources, remains a pressing issue.