Experimental study of the efficiency of scalable YOLOv8(n-x) architectures in multi-object tracking tasks based on the ByteTrack algorithm

Abstract

The article is devoted to the results of research on the performance of a computer vision system for intelligent road traffic monitoring tasks. The problem of finding a balance between Inference Speed and Tracking Stability in dense urban traffic conditions is investigated.

Classic approaches to multi-object tracking (MOT) often face the problem of object loss during partial overlap (occlusion). The modern ByteTrack algorithm offers a solution to this problem by using two-level association of detections with existing tracks, which makes it particularly resistant to overlaps. This algorithm is the optimal solution for real-time systems, as it minimises computational costs while maintaining the highest trajectory retention quality among the considered analogues. The basic principle of the IoU tracker, which underlies the geometric component of ByteTrack, was considered. The paper experimentally proves that for real-time urban monitoring tasks on mid-range hardware (NVIDIA RTX 2060 level), the YOLOv8m model is the best choice. The paper provides a detailed analysis of the five tested YOLOv8m architectures: Nano, Small, Medium, Large, and X-Large. Through empirical comparison of the performance and quality of models of different scales based on real video data from a traffic intensity monitoring system, the optimal neural network architecture of the YOLOv8 family for integration with ByteTrack was determined.

As a result of the analysis of modifications of the YOLOv8 neural network in combination with the ByteTrack tracker, a conclusion is drawn regarding a significant non-linear dependence between the computational complexity of the neural network detector and the final quality of multi-object tracking.