Gated box jittering and data augmentation scheme optimized soft-teacher for semi-supervised object detection

Abstract Auto-driving trains are essential in today's intelligent transportation, with object detection being key in automated systems. This study presents the Nanchang Railway Group Large-Scale Dataset to fill the gap in data with limited temporal variation. It introduces a novel semi-supervised object detection method, the gated box-jittering soft-teacher (GBJST), which addresses labour costs in data processing and improves detection accuracy in dense object scenarios. The research demonstrates that combining reduced weak data augmentation with enhanced strong augmentation significantly improves detection accuracy, especially with limited labelled data. The GBJST method showed an average improvement in precision, confirming its effectiveness. By refining the data augmentation approach and incorporating the GBJST, this method realized an average improvement of 1.96%, 2.03% and 2.86% in mean average precision (mAP) on datasets with 10% labelled data, thereby affirming the algorithm's feasibility.