Abstract
Autonomous vehicles (AVs) require precise, real-time Field of View (FOV) estimation to distinguish between visible and occluded regions. Traditional geometric algorithms are highly vulnerable to adversarial LiDAR spoofing, which can manipulate a vehicle's perception. This paper introduces a deep learning-based approach that treats FOV estimation as a binary segmentation task. By employing a UNet architecture with Monte Carlo Dropout (MCD) for uncertainty quantification, the proposed system provides robust visibility estimates and real-time anomaly detection for sensor attacks. Evaluated on simulated and real-world datasets, the model achieves F1-scores above 0.85 and operates at 50 Hz, demonstrating its suitability for secure autonomous navigation.
Introduction
Security-aware sensor fusion is a cornerstone of autonomous vehicle safety. It involves verifying sensor data against environmental expectations. A key part of this verification is knowing the vehicle's Field of View (FOV)—the specific areas currently visible to sensors, accounting for occlusions from buildings, vehicles, and terrain.
Historically, FOV estimation has relied on classical computer graphics techniques such as ray tracing or concave hull estimation. However, these geometric methods possess a critical security flaw: they "blindly trust" incoming data. If an attacker injects fake LiDAR points (spoofing), these algorithms expand the estimated FOV to include the malicious data. This allows attackers to create "ghost" obstacles or hide real ones, potentially leading to catastrophic failures.
Methodology
The authors propose a learning-based framework that shifts FOV estimation from a purely geometric problem to a probabilistic segmentation task.
1. Data Representation
3D LiDAR point clouds are first projected into a 2D Bird’s-Eye-View (BEV) representation. This involves filtering out irrelevant points (e.g., treetops or ground planes) and quantizing the remaining data into a structured grid. This 2D format allows for efficient processing using standard image segmentation architectures.
2. Neural Network Architecture
The core of the system is a UNet architecture, which is well-suited for dense pixel-wise (or cell-wise) predictions. The model is trained to predict the probability of visibility for each grid cell in the BEV map. Unlike geometric algorithms, the DNN learns environmental "priors" — it understands the typical patterns of LiDAR returns and can distinguish realistic visibility from anomalous patterns.
3. Uncertainty and Security
To handle adversarial environments, the authors integrate Monte Carlo Dropout (MCD). During inference, multiple forward passes are performed with stochastic dropout enabled. This generates a "confidence map" alongside the visibility prediction. When the model encounters spoofed points that deviate from learned LiDAR patterns, the variance in its predictions increases significantly. This high uncertainty serves as a real-time alarm for detecting sensor attacks.
Datasets and Evaluation
A significant hurdle in FOV research is the lack of ground-truth visibility labels in standard autonomous driving datasets. To address this, the authors developed a labeling pipeline and released three distinct datasets: CARLA (simulated), nuScenes (real-world outdoor), and UGV (real-world indoor).
Key Results
The proposed segmentation model significantly outperforms traditional methods in both accuracy and security:
- Attack Robustness: Under LiDAR spoofing attacks, the precision of traditional ray tracing and concave hull methods drops to nearly 0%. The proposed model maintains high precision and recall by effectively ignoring points that do not conform to learned environmental priors.
- High Performance: Across all three datasets, the models achieved F1-scores typically exceeding 0.85.
- Real-Time Efficiency: The entire pipeline, including preprocessing and inference, operates at over 50 Hz. This low latency is crucial for deployment on AV hardware.
- Generalization: The models demonstrated impressive "zero-shot" transferability, with models trained on simulated CARLA data performing reliably on real-world nuScenes data.
Conclusion and Future Work
This work demonstrates that probabilistic segmentation is a superior alternative to geometric FOV estimation for secure autonomous systems. By leveraging learned priors and uncertainty quantification, the system provides a robust defense against LiDAR spoofing. Future research will focus on developing optimal "recovery" strategies once an attack is detected and exploring 3D volumetric representations to better handle vertically complex environments.
Summary updated