Anomaly Detection in Complex Driving Scenes for Autonomous Driving

Abstract

Autonomous driving technology represents a significant shift in transportation, promising improved safety, efficiency, and a new level of convenience. However, the full benefits and the adoption of this technology depend on the effectiveness of anomaly detection systems in autonomous vehicles. Anomalies are deviations from normal driving patterns that can pose serious safety risks and must be detected and managed in real-time.This thesis highlights the significance of anomaly detection for autonomous vehicles and investigates approaches that improve the perception capabilities of autonomous systems. Anomaly detection is critical in strengthening the safety features of autonomous driving systems by accurately identifying and managing a range of unpredictable scenarios, such as adverse weather conditions, sudden pedestrian movements, or abnormal road conditions. This advancement is vital in ensuring that autonomous driving systems are dependable, robust, and capable of handling various challenging situations. The success of autonomous systems relies on their ability to accurately perceive and interpret complex environments. Therefore, this study focuses on exploring scene understanding in autonomous driving. In this work, we investigate the latest literature that categorizes anomalies for camera data and provide an in-depth analysis of recent anomaly detection datasets. This analysis shows the evolving complexity and variability of anomalies that autonomous systems must address to maintain the safety of autonomous vehicles. This work investigates the state-of-the-art models for detecting anomalies in autonomous driving, which is crucial for identifying unusual or hazardous objects on the road. Our study involves training an anomaly segmentation framework and evaluating its performance in complex driving environments, ranging from rural to urban conditions.