Framework for Hazardous Situations detection in Autonomous Driving

Abstract

The advancement of Autonomous Vehicles (AVs) and Advanced Driver Assistance Systems (ADAS) hinges on their ability to perceive and respond to complex, dynamic environments, particularly in rare and safety-critical scenarios. While current systems excel under nominal conditions, detecting and reacting to anomalous events –such as sudden pedestrian appearances or erratic driver behavior– remains a significant challenge due to the scarcity of such events in training datasets. This paper presents a modular, semi-supervised approach for detecting risk-driving situations using monocular front-camera imaging. The proposed two-stage framework decouples visual perception and anomaly detection: the first stage employs instance segmentation algorithm YOLO to extract scene features; while the second stage identifies anomalies using these features, trained solely on non-risk scenarios using one-class Support Vector Machine. This design enables independent optimization of each module and supports the integration of both classical machine learning and deep learning techniques for anomaly detection. Proposed methodology has been tested with the nuScenes dataset, reaching a 0.31 recall and 0.70 accuracy, demonstrating the capacity of hybrid algorithms combining computer-vision with machine-learning to identify hazardous driving situations while maintaining a lightweight execution environment.