Unsupervised Hybrid VAE-Based Anomaly Detection for Vehicle Onboard LiDAR Sensors
Intelligent transportation systems (ITS) are revolutionizing road safety, particularly in urban areas. Innovative sensors like LiDAR are being deployed to monitor traffic flow in real-time, providing precise data on vehicle movements, road conditions, and congestion patterns. These advancements open...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10994438/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Intelligent transportation systems (ITS) are revolutionizing road safety, particularly in urban areas. Innovative sensors like LiDAR are being deployed to monitor traffic flow in real-time, providing precise data on vehicle movements, road conditions, and congestion patterns. These advancements open the path for safer roads and more efficient transportation, with the potential for autonomous driving in the future. In this context, since various environmental factors, such as weather fluctuations and terrain diversity, can introduce anomalies that pose risks and malfunctions to a variety of sensor driving systems, the detection of anomalies in spatial-temporal (ST) preprocessed LiDAR data becomes crucial to ensuring safety. In this paper, we propose a novel low-complex unsupervised model for anomaly detection (AD) within ST preprocessed LiDAR data named CNN-BiLSTM VAE that combines variational auto-encoder (VAE) reconstruction capabilities, convolutional neural networks (CNN) spatial characteristic learning capabilities, and bidirectional long-short-term memory (BiLSTM) networks time series learning capabilities in a symmetric mirror-to-mirror (M2M) architecture. Our method aims to detect abnormal behavior and unusual ST patterns in preprocessed, unlabeled LiDAR data without any prior knowledge of the presence of anomalies by learning complex patterns and dependencies within the ST data, while considering the effects of multiple environmental factors. Experimental validation against benchmark models demonstrates the effectiveness of our proposed model in detecting anomalies within preprocessed multidimensional ST LiDAR data. The model strikes a good balance between complexity and efficiency, achieving a 10% accuracy improvement compared to existing models, with an accuracy of 95.1% and an F1-score of 82.6%. Additionally, injecting diverse environmental anomalies enables testing the model’s performance and reliability in real-world scenarios influenced by environmental factors, ensuring robustness across various conditions. |
|---|---|
| ISSN: | 2169-3536 |