A Loss Function with Periodicity-Aware Scoring for Crosswalk Detection

Abstract

We propose a novel loss function incorporating a periodicity-aware scoring method to improve crosswalk detection. This study aims to develop a driver support system using dashcam footage, where crosswalk detection plays a crucial role as crosswalks are key elements in driving scenes. However, conventional object detection models often fail to detect crosswalks accurately. To address this limitation, we focus on the alternating brightness patterns that characterize crosswalks. By integrating a periodicity-aware scoring mechanism into the loss function, this approach significantly improves the model’s ability to detect patterns like crosswalks.