YOLOv11 for Robust Traffic Monitoring: Enhanced Detection of Minority Vehicle Classes

Abstract

This study aims to evaluate the effectiveness of the YOLOv11 segmentation instance model and also to address the problem of managing the unbalanced dataset obtained from traffic data in Makassar City, Indonesia, which is related to the vehicle counting system. The main objective of this study is to test how effective the weighted-class balance YOLOv11 model is in detecting vehicles, especially in traffic congestion conditions. The proposed method dynamically assigns higher weights to minority classes during training, eliminating the need to add files directly to the dataset for oversampling. Performance was evaluated using mAP 50, mAP 50-95, precision, recall, and F1-score, with statistical validation conducted via the Wilcoxon signed-rank test, Friedman test, and Cohen’s d effect size. The results showed YOLOv11 (weighted) outperformed other models, both YOLOv8 (weighted) and YOLOv11(baseline), resulting in mAP 50-95 scores of 0.717 (mask) and 0.836 (bounding box), with an improvement of 27.8 percent for tricycles (bounding box) and 26.5 percent for buses (mask) over baseline values. Statistical tests confirmed the significant improvement ($\mathbf{p}=\mathbf{0 . 0 3 1}$ for both tasks) and large effect size ($\mathbf{d} \boldsymbol{\gt} \mathbf{1 . 3}$ for mask and $\mathbf{d} \boldsymbol{\gt} \mathbf{1 . 4}$ for bounding box), indicating statistical and practical relevance. The method also maintained robustness to the majority class while overcoming the challenge of overlapping objects. These findings validate YOLOv11 (weighted) as a scalable solution for traffic management systems in congested environments.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint