Event-Based Visual Odometry on Non-Holonomic Ground Vehicles

Abstract: Despite the promise of superior performance under challenging conditions, event-based motion estimation remains a hard problem owing to the difficulty of extracting and tracking stable features from event streams. In order to robustify the estimation, it is generally believed that fusion with other sensors is a requirement. In this work, we demonstrate reliable, purely event-based visual odometry on planar ground vehicles by employing the constrained non-holonomic motion model of Ackermann steering platforms. We extend single feature n-linearities for regular frame-based cameras to the case of quasi time-continuous event-tracks, and achieve a polynomial form via variable degree Taylor expansions. Robust averaging over multiple event tracks is simply achieved via histogram voting. As demonstrated on both simulated and real data, our algorithm achieves accurate and robust estimates of the vehicle's instantaneous rotational velocity, and thus results that are comparable to the delta rotations obtained by frame-based sensors under normal conditions. We furthermore significantly outperform the more traditional alternatives in challenging illumination scenarios. The code is available at \url{https://github.com/gowanting/NHEVO}.

• The paper introduces a purely event-based visual odometry system leveraging non-holonomic Ackermann models for robust motion estimation.
• It employs polynomial formulations and Taylor series expansions to adapt traditional n-linear constraints to asynchronous event data.
• Histogram voting and rank minimization techniques enhance feature tracking accuracy under challenging lighting conditions.

Event-Based Visual Odometry on Non-Holonomic Ground Vehicles

The study by Xu et al. addresses the complexities inherent in achieving effective visual odometry (VO) using event-based cameras on non-holonomic ground vehicles. Traditional visual odometry techniques using regular cameras face significant performance hurdles in high-dynamic and high-dynamic-range conditions due to motion blur and overexposure. Event-based cameras, which capture changes in illumination asynchronously on a pixel-by-pixel basis, offer a promising alternative by providing high dynamic range and rapid response times. However, processing event streams for stable feature extraction and tracking poses substantial challenges.

The authors propose a purely event-based visual odometry solution by leveraging the constraints of non-holonomic motion specific to Ackermann steering models, conventions commonly used in vehicle dynamics. The adaptation of the non-holonomic ground vehicle kinematic model into VO tasks allows for more robust motion estimation. The study extends single feature n-linearities, typically used in traditional frame-based visual systems, to accommodate the quasi time-continuous dynamics captured through event cameras. The authors employ polynomial formulations via variable Taylor expansions to create a framework that accommodates dynamic, event-driven data streams.

The paper highlights an advanced approach of averaging across multiple event tracks using histogram voting, which improves the resilience of odometry results. This method shows improved performance on vehicle’s instantaneous rotational velocity estimation, demonstrated via extensive simulations and real-world data evaluations. Notably, the proposed method exhibits superior performance under challenging lighting conditions compared to traditional frame-based methods, confirming its potential for applications in environments where regular cameras are less effective.

Key contributions of the paper include:

• Integration of a continuous-time incidence relation within the VO framework under the assumption of non-holonomic Ackermann motion and constant rotational velocity.
• Utilization of Taylor series expansions to approximate trigonometric functions, generating n-linear constraints related to rotational velocity.
• Deployment of rank minimization and histogram voting techniques for robust solution extraction and outlier rejection.

The implications of this research are both practical and theoretical. Practically, it provides a more reliable method for visual navigation in automated ground vehicles operating in challenging environments, potentially improving the efficacy and safety of autonomous vehicles. Theoretically, it broadens the application scope of event-based cameras within robotics, encouraging further study into non-holonomic constraints in diverse robotic and automotive applications.

Future research may explore the integration of more advanced sensor fusion techniques, such as incorporating inertial measurements, to further enhance the robustness and accuracy of estimation processes in motion prediction frameworks. Additionally, advancements in event-camera hardware and asynchronous signal processing could further bridge the performance gaps with traditional visual systems, unlocking new potential in real-time navigation and mapping tasks.