TL;DR
This paper introduces a novel method for estimating a mobile robot's velocity using event-based optical flow (eOF) from a camera pointed at the ground. This approach decouples velocity estimation from wheel odometry, offering improved accuracy, especially for lateral velocity, and robustness to slippery surfaces, outperforming state-of-the-art methods in challenging scenarios.
What Problem Does This Paper Address?
Accurate velocity estimation is crucial for autonomous driving and advanced driver-assistance systems (ADAS). Traditional methods, like wheel odometry fused with IMU data, often rely on assumptions of non-slip steering or complex vehicle dynamics models that fail in real-world conditions, particularly on slippery surfaces. These methods also struggle with estimating lateral velocity. While visual-inertial odometry (VIO) offers an alternative, it can be sensitive to initialization and environmental conditions. The paper addresses the need for a robust, accurate, and decoupled velocity estimation method that can handle fast-moving vehicles and challenging terrains, specifically focusing on improving lateral velocity estimation.
Methods
The core of the proposed method is an event-based optical flow (eOF) algorithm. Event cameras, unlike traditional frame-based cameras, capture intensity changes asynchronously with high dynamic range and low latency, making them resistant to motion blur.
- Data Preprocessing: Raw event data is accumulated into 2D histograms over a defined accumulation time, creating an image-like representation. This accumulation time acts analogously to exposure time in conventional cameras.
- Optical Flow Computation: The Farneback algorithm, a classical optical flow method, is applied to these event histograms to compute dense, per-pixel optical flow vectors.
- Rigid Body Motion Estimation: Assuming the camera is mounted facing the ground and observing a planar scene, the problem is simplified to estimating a 2D rigid body motion (translation and rotation). The computed optical flow vectors are used to solve for this motion using singular value decomposition.
- Outlier Removal: A RANSAC (Random Sample Consensus) loop is incorporated to robustly estimate motion by iteratively selecting samples of flow vectors and identifying inliers, mitigating the impact of erroneous flow calculations.
- Rigid Body Transformation: The estimated camera velocity is transformed to the vehicle's rear axle, accounting for the camera's position relative to the axle.
- Augmentation with IMU Data: To improve yaw rate estimation, which is critical for accurate lateral velocity calculation, IMU data is integrated. The paper investigates replacing the yaw rate estimated from eOF with raw IMU yaw rate measurements.
Datasets
The paper utilizes the DSEC-Flow dataset for evaluating the performance of different optical flow algorithms. For their own method's evaluation, they conducted in-field experiments using a 1:10 scale autonomous racing platform equipped with a DAVIS 346 event camera and a Vicon motion-capture system for ground truth. They also performed full-scale highway experiments using a Prophesee EVK4 event camera mounted on a real car.
Key Insights and Ideas
- Event cameras are ideal for ground-facing velocity estimation: Their low latency and high dynamic range effectively combat motion blur, a significant issue for conventional cameras when observing fast-moving scenes close to the ground.
- Planar kinematics simplifies the problem: By assuming the camera observes a rigid body moving in a plane parallel to the ground, the complex 3D motion estimation is reduced to a simpler 2D rigid body transformation problem.
- Direct optical flow is competitive: The proposed eOF method, by directly estimating velocity from optical flow, achieves performance comparable to, and in some aspects superior to, more complex Visual-Inertial Odometry (VIO) methods, particularly in lateral velocity estimation.
- IMU fusion significantly improves yaw rate estimation: Integrating IMU data directly enhances the accuracy of the yaw rate, which in turn leads to a substantial improvement in lateral velocity estimation.
Key Results
- Quantitative Improvement: The eOF method achieved a 6.5% improvement in longitudinal velocity RMSE and a 38.3% improvement in lateral velocity RMSE compared to the forward-facing eVIO (Ultimate SLAM) method.
- Lateral Velocity Accuracy: The eOF method with IMU correction achieved a 41.1% reduction in lateral velocity RMSE compared to the pure eOF implementation.
- Highway Performance: In full-scale highway experiments at speeds up to 32 m/s, the method demonstrated robust velocity estimation, with a mean estimated velocity of 31.28 m/s (2.3% error) without RANSAC, and 31.88 m/s (0.4% error) with RANSAC, showing a standard deviation of 0.64 m/s.
- Computational Efficiency: The eOF algorithm exhibits a comparable CPU load and memory usage to eVIO, with a slightly higher latency (60.35 ms vs. 1.04 ms for eVIO), though direct comparison is nuanced due to asynchronous nature of eVIO.
Limitations and Future Work
- Featureless or Flat Surfaces: The method's reliance on optical flow means it can struggle in environments with very few features or completely flat surfaces.
- Scale Adjustment: For precise metric velocity estimation, scale adjustment using a laser distance sensor is advised to account for variations in camera height due to suspension.
- Future Work: The authors suggest training a dedicated optical flow network for event cameras to explore different event representations and potentially achieve even better flow estimates. Further validation on additional real-world data is also recommended.
Summary updated