Artificial Intelligence in Industry 4.0 and 5G Technology

15Re‐Identification‐Based Models for Multiple Object Tracking

Alexey D. Grigorev¹, Alexander N. Gneushev^1,2, and Igor S. Litvinchev²

¹Moscow Institute of Physics and Technology, Department of Control and Applied Mathematics, Dolgoprudny, Moscow Region, Russia

²Federal Research Center “Computer Science and Control” of Russian Academy of Sciences, Dorodnicyn Computing Center, Moscow, Russia

15.1 Introduction

The problem of multiple object tracking (MOT) is very relevant in the field of computer vision due to video surveillance systems that have developed a lot over the past decade. Tracking tasks arise both in security systems, automation control, self‐driving, and robotics. Online processing of incoming data provides an opportunity to reduce the computational load on the system and process incoming data locally, regardless of the availability of a remote server.

MOT is associated with objects' detection in a sequence of video frames that are an integral part of systems for identifying and verifying objects, assessing their spatial position. Special detection models are used for coarse localization of objects in the image; however, these detectors require significant computational resources [1–3]. Because many objects are presented in the image, tracking models help clarify their coordinates by determining a separate track of each object's movement by measuring its position in each frame and estimating movement in future moments in time. The predicted object position can significantly ...