Out-of-Distribution Object Detection in Autonomous Vehicles With Yolo Model

This paper addresses the challenge of detecting Out-of-Distribution (OOD) objects in autonomous vehicles, focusing on identifying and localizing objects absent from the training data. We propose a novel method that leverages the existing object detection model, YOLOv5, to detect OOD instances without requiring model retraining or additional datasets. Our approach computes dissimilarity scores from class confidence outputs to effectively distinguish OOD objects in cropped images. Experiments on popular autonomous vehicle 2D object detection datasets demonstrate that, in a straightforward scenario, our method significantly reduces the False Positive Rate at 95% True Positive Rate while maintaining a comparable Area Under the Receiver Operating Characteristic curve (AUROC) to baseline models. In more challenging scenarios, our method outperforms competitors, demonstrating superior robustness. Key contributions include the proposed OOD detection method and the methodology for identifying OOD object instances