Few-shot learning for novel object detection in autonomous driving

Artificial intelligence and advanced sensing technologies have significantly advanced the intelligent transportation system and autonomous vehicles. Perception, a critical component, extracts real-time traffic information essential for various system functionalities, such as agent behavior prediction. However, the quality of information derived from perception greatly influences overall system performance. This study focuses on enhancing perception robustness in autonomous vehicles, particularly in detecting rare objects, which pose a challenge due to limited training samples. While deep learning-based vision methods have shown promising accuracy, they struggle with rare object detection. To address this, we propose a few-shot learning training strategy tailored for improved detection accuracy of rare or novel objects. Additionally, we design a one-stage object detector for efficient object detection in autonomous driving scenarios. Experiments on a self-driving dataset augmented with rare objects alongside the popular few-shot object detection (FSOD) benchmark, the pattern analysis, statical modeling, and computational learning PASCAL Visual Object Classes (PASCAL-VOC), demonstrate state-of-the-art accuracy in rare categories and superior inference speed compared to alternative algorithms. Furthermore, we investigate the impact of intra-class variance on detection accuracy, providing insights for data annotation in the preparation stage.