Planar Velocity Estimation for Fast-Moving Mobile Robots Using Event-Based Optical Flow

TL;DR

The paper tackles robust velocity estimation for fast-moving mobile robots without relying on wheel-ground traction models by leveraging planar kinematics and ground-facing event-camera optical flow. It introduces a simple, direct estimator that converts ground-plane optical flow into rear-axle velocity through a 2D rigid-body motion model, with RANSAC outlier rejection and optional IMU yaw-rate augmentation. Quantitative results show performance on par with state-of-the-art Event-VIO, including a 38.3% improvement in lateral velocity error, and qualitative highway-speed experiments up to 32 m/s demonstrate real-world viability. The approach offers a low-latency, high-dynamic-range alternative to frame-based methods, and an open-source implementation pathway is noted for broader adoption.

Abstract

Accurate velocity estimation is critical in mobile robotics, particularly for driver assistance systems and autonomous driving. Wheel odometry fused with Inertial Measurement Unit (IMU) data is a widely used method for velocity estimation; however, it typically requires strong assumptions, such as non-slip steering, or complex vehicle dynamics models that do not hold under varying environmental conditions like slippery surfaces. We introduce an approach to velocity estimation that is decoupled from wheel-to-surface traction assumptions by leveraging planar kinematics in combination with optical flow from event cameras pointed perpendicularly at the ground. The asynchronous micro-second latency and high dynamic range of event cameras make them highly robust to motion blur, a common challenge in vision-based perception techniques for autonomous driving. The proposed method is evaluated through in-field experiments on a 1:10 scale autonomous racing platform and compared to precise motion capture data, demonstrating not only performance on par with the state-of-the-art Event-VIO method but also a 38.3 % improvement in lateral error. Qualitative experiments at highway speeds of up to 32 m/s further confirm the effectiveness of our approach, indicating significant potential for real-world deployment.

Figures (6)

• Figure 1: Overview of the proposed method. Depicted on the left is the experimental setup used during the full-scale highway experiment as well as an example event frame captured at 32ms and accumulated over 100µs.
• Figure 2: Comparison of the maximum allowable camera exposure time to not exceed the indicated levels of motion blur at different vehicle speeds. A fixed camera height of 0.6m and of 60° are assumed.
• Figure 3: Qualitative visualization of a conventional camera image (left) at 200 and an accumulated event camera image (right) using an accumulation time of 1m s (1000Hz). The disk positioned at 0.3m from the cameras with a radius of 0.15m is spinning at 360.0 for both cameras, resulting in a rotational velocity of 5.65m s at the outer edge of the disk. The recordings were taken at 387Lux, with an exposure time of 0.87m s for the conventional camera. vectors depicted as arrows (orange), pixel-level brightness changes of the event camera visualised in (blue/black).
• Figure 4: A depiction of the velocity estimation performance in-field experiment. On the left, a 1:10 scaled autonomous racing car is depicted with the DAVIS 346 event camera mounted. On the right, the track is set within a motion-capturing setup. The red circles highlight the motion-capture tracking components, i.e., reflective markers and cameras.
• Figure 5: Qualitative visualization of the scaled autonomous driving velocity experiment with motion-capture ground truth comparison. On the left, the tracked trajectory is depicted in red with track advancement $s$ denoted every 2m. On the right the $v_{lon}, v_{lat}, v_{\psi}$ velocity states of front-facing (blue), downwards-facing (green), and motion-capture ground truth measurements (red). The vertical red dashed lines represent the 0m mark of the track depicted on the left.