Comprehensive Analysis of YOLOv8 + DeepSORT for Vehicle Tracking: HOTA and CLEAR-Based Evaluation

Abstract

This paper offers a thorough comparative investigation of the performance of a vehicle multi-object tracking system, incorporating various versions of the YOLOv8 detector (from ‘n’ to ‘x’) alongside the DeepSORT tracking algorithm. This study systematically assesses the impact of the trade-off between detector speed and accuracy on tracking metrics, utilising a real-world traffic video dataset from Bali. The assessment is performed utilising two fundamentally distinct metric frameworks: the traditional CLEAR metric (which includes MOTA) and the contemporary Higher Order Tracking Accuracy (HOTA) metric. The findings indicate that although the larger YOLOv8 model markedly enhances detection recall, particularly for smaller and more difficult items like motorcycles, tracking issues persist. The dual metric study provides significant insights: the HOTA measure demonstrates that car tracking has more associative stability (higher AssA scores) compared to motorbike tracking, which frequently experiences track fragmentation. In contrast, the detection-biased MOTA metric produces somewhat paradoxical outcomes, as motorbikes receive elevated scores due to enhanced detection accuracy (fewer false positives), therefore obscuring deficiencies in tracking consistency. This study concludes that HOTA offers a more comprehensive evaluation by differentiating between detection and association performance, so demonstrating that detection-only metrics like MOTA can yield an imperfect representation of actual tracking ability. These findings underscore the necessity of matching detector architecture and evaluation criteria with specific application requirements, particularly in safety-critical systems where identity consistency is essential.