Hybrid Video Anomaly Detection for Anomalous Scenarios in Autonomous Driving

In autonomous driving, the detection of anomalies is crucial for ensuring safety and reliability. Video anomaly detection (VAD) focuses on identifying events in video data that deviate from an expected normality. In autonomous driving, the challenges of detection are complicated by factors such as camera movements, ever-changing backgrounds, and rapid changes in vehicle speed. Many different types of anomalies exist [1, 2, 3], with many approaches trying to detect them [4]. We can distinguish between five key techniques: Reconstruction, prediction, generative modeling, feature extraction, and confidence evaluation [5].