Performance Comparison of Semantic Segmentation Models and Loss Functions for Seat Belt Detection

Abstract

In Japan, despite seat belts being mandatory for drivers, some choose not to comply. This non-compliance is typically identified through visual inspections by police officers. However, Japan's population decline has facilitated a growing need for automating this process using camera and vision technologies. This study explores the optimal combination of semantic segmentation models and loss functions for seat belt detection in images. We created a dataset using various car models in outdoor settings and evaluated the performance of all combinations of nine models and five loss functions. Our findings indicate that the UNet++ model paired with the Lovász loss function delivers superior performance.