Application of machine learning models for tracking and counting heavy vehicles on highways

Accurate traffic volume estimation is essential for road infrastructure planning, particularly for the design and maintenance of asphalt pavements. This study presents a comparative evaluation of two object tracking algorithms, BoT-SORT and ByteTrack, applied to the task of monitoring heavy vehicles in real-world road environments using side-view videos. Detection was performed using both pre-trained and customized YOLOv8 models, with the latter tailored to the vehicle categories defined by the DNIT traffic manual. A single-line counting technique was employed to mitigate false negatives caused by occlusions. The experiments were conducted using real data from the BR-110 highway, and performance was evaluated based on precision, recall, and F1 score metrics. The results demonstrate that customized YOLOv8 models, particularly when combined with ByteTrack, achieved superior performance, with a significant reduction in false positives. Previous studies indicate that camera positioning and model adaptation to specific data are critical factors for enhancing detection accuracy in side-view scenarios, especially in environments subject to occlusions and heavy traffic. In this study, we observed that the use of customized models led to improved performance; however, no comparative experiments were conducted with different camera positions, which limits the generalizability of the findings.