FlowCalib: LiDAR-to-Vehicle Miscalibration Detection using Scene Flows

TL;DR

FlowCalib addresses the problem of LiDAR-to-vehicle miscalibration by leveraging scene-flow patterns of static objects, eliminating the need for extra sensors. It combines a neural scene-flow prior for flow estimation with a dual-branch detector that fuses global flow embeddings and handcrafted geometric descriptors to perform global and axis-specific miscalibration detection. The method uses fault injection, robust preprocessing, NSFP-based scene flow generation, and a two-headed detector to output miscalibration indicators, achieving strong performance on nuScenes and providing interpretable axis-level insights. This work has practical impact by enabling sensor attribution and reliable miscalibration detection in autonomous driving without calibration targets or additional sensing modalities.

Abstract

Accurate sensor-to-vehicle calibration is essential for safe autonomous driving. Angular misalignments of LiDAR sensors can lead to safety-critical issues during autonomous operation. However, current methods primarily focus on correcting sensor-to-sensor errors without considering the miscalibration of individual sensors that cause these errors in the first place. We introduce FlowCalib, the first framework that detects LiDAR-to-vehicle miscalibration using motion cues from the scene flow of static objects. Our approach leverages the systematic bias induced by rotational misalignment in the flow field generated from sequential 3D point clouds, eliminating the need for additional sensors. The architecture integrates a neural scene flow prior for flow estimation and incorporates a dual-branch detection network that fuses learned global flow features with handcrafted geometric descriptors. These combined representations allow the system to perform two complementary binary classification tasks: a global binary decision indicating whether misalignment is present and separate, axis-specific binary decisions indicating whether each rotational axis is misaligned. Experiments on the nuScenes dataset demonstrate FlowCalib's ability to robustly detect miscalibration, establishing a benchmark for sensor-to-vehicle miscalibration detection.

Figures (6)

• Figure 1: This is a qualitative illustration of objects in the environment model in the presence of sensor-to-vehicle yaw angle errors. Objects appearing in neighboring lanes can pose safety issues during autonomous vehicle operation. $\{V\}$ represents the vehicle frame and $\{L\}$ the LiDAR frame.
• Figure 2: The plots on the left show the performance and sensor alignment in the presence of a yaw angle LiDAR-to-vehicle error. The right plots show the performance and LiDAR-camera alignment in the presence of an inverse yaw angle camera-to-vehicle error. From the LiDAR-to-camera perspective, the two sensors have identical errors. However, significant performance drops are observed only for LiDAR-to-vehicle errors.
• Figure 3: The idea behind using flow fields to detect LiDAR miscalibration is that a calibration error appears as a bias in the point flow. This causes distance-dependent errors that decrease as an object approaches the vehicle. The vehicle's trajectory is shown above. The resulting motion pattern of the objects, represented here as points, is shown below.
• Figure 4: FlowCalib uses a two-stage learning process. First, flow fields are generated to construct features. These features are used to learn feature embeddings and train the global and axis detection heads. The heads are trained to detect miscalibration in the presence of random calibration errors and to identify the miscalibrated axis.
• Figure 5: The point cloud is shown before (left) and after (right) the pre-processing steps. The blue and green point clouds after preprocessing are shown at timesteps $t$ and $t+1$, respectively. These serve as input for scene flow generation. Note that the point cloud are transformed into vehicle coordinate systems after the preprocessing.