Anomaly Detection in Autonomous Vehicle’s Lidar Sensor Data Using Variational Autoencoders

LiDAR sensor data is essential for autonomous vehicle navigation, traffic flow monitoring, obstacle detection, and passenger safety. However, the reliability of LiDAR data can be compromised by anomalies caused by sensor malfunctions, environmental conditions, or unexpected road events. To address this, detecting anomalies in spatial-temporal (ST) LiDAR data is critical for ensuring safety. This paper proposes a novel low-complexity unsupervised framework named CNN-BiLSTM VAE for anomaly detection (AD) in non-image LiDAR data. The framework combines variational auto-encoder (VAE) reconstruction, CNN for spatial learning, and bidirectional LSTM for time-series learning in a mirror-to-mirror (M2M) architecture. Experimental results show that this method effectively detects anomalies in multidimensional ST LiDAR data, thereby maintaining robustness under various environmental conditions.