Re-identification problem for objects with independent trajectories

Abstract

Multi-Object Tracking (MOT) systems represent one of the key areas of modern computer vision and artificial intelligence, aimed at detecting and tracking multiple moving objects in real time. These systems not only identify objects but also maintain their identities across sequential frames over time. However, the accuracy and reliability of multi-object tracking systems are often challenged by various technical issues, such as object occlusion—when one object is temporarily hidden by another object or by the background—which significantly reduces tracking precision. In such situations, the system may lose track of the object or mistakenly associate it with another one. Furthermore, factors like objects temporarily leaving the camera’s field of view, variations in lighting, movement speed, and background complexity can also negatively affect tracking performance. To overcome these challenges, modern MOT systems employ advanced algorithms such as deep learning, convolutional neural networks (CNNs), recurrent neural networks (RNNs), transformer architectures, and object re-identification (Re-ID) technologies. These approaches are particularly effective in predicting object trajectories and detecting or restoring temporarily lost objects, thereby improving the overall accuracy and robustness of tracking systems