Balancing Helmet Detection with Synthetic Data for Class Imbalance

초록

Detecting helmet compliance among motorcyclists is essential to mitigate head injuries. However, training robust vision models for this task is hindered by a significant class imbalance: images of helmeted riders are abundant, whereas those of non-helmeted riders are scarce. This imbalance degrades the model's ability to accurately detect safety violations. Our initial experiments showed that naively augmenting the minority class (non-helmeted riders) with synthetic images without adjusting class proportions degrades detection performance. We hypothesized that class ratio alignment - carefully balancing the distribution of helmeted vs. non-helmeted images in the dataset through the inclusion of synthetic data - is essential for performance improvement. We propose a ratio-aware augmentation method that synthesizes non-helmeted rider images while maintaining class balance and demonstrate that this strategy significantly improves helmet violation detection accuracy. These findings highlight the importance of pairing synthetic augmentation with class distribution calibration. © 2025 IEEE.

키워드