VibES: Induced Vibration for Persistent Event-Based Sensing

TL;DR

VibES addresses the core limitation of event cameras in static scenes by mechanically inducing motion with a rotating unbalanced mass to sustain event generation. A model-based, real-time motion-compensation pipeline using an EKF and NUFFT estimates recovers motion-free scene information, enabling improved image reconstruction and edge detection. The approach is validated on a hardware prototype with real and synthetic datasets, showing higher information content, cleaner edges, and robust downstream utility, including frequency estimation and relative depth cues. This hardware-simple, software-centric method has potential for real-time, low-power perception in robotic systems and can be extended with adaptive vibration control and broader integration with event-based pipelines.

Abstract

Event cameras are a bio-inspired class of sensors that asynchronously measure per-pixel intensity changes. Under fixed illumination conditions in static or low-motion scenes, rigidly mounted event cameras are unable to generate any events and become unsuitable for most computer vision tasks. To address this limitation, recent work has investigated motion-induced event stimulation, which often requires complex hardware or additional optical components. In contrast, we introduce a lightweight approach to sustain persistent event generation by employing a simple rotating unbalanced mass to induce periodic vibrational motion. This is combined with a motion-compensation pipeline that removes the injected motion and yields clean, motion-corrected events for downstream perception tasks. We develop a hardware prototype to demonstrate our approach and evaluate it on real-world datasets. Our method reliably recovers motion parameters and improves both image reconstruction and edge detection compared to event-based sensing without motion induction.

Figures (12)

• Figure 1: Qualitative illustration of our method.(a) No Vibration. With a static event camera and no induced motion, the accumulated event image appears blurred and lacks sharp edges, while the $y$–$t$ slice shows little temporal structure. (b) With Vibration. Introducing controlled vibrations stimulates the sensor, increasing the number of events and producing sinusoidal traces in the $y$–$t$ slice. We develop an Extended Kalman Filter (EKF) to track the vibrational motion in a region of the scene (red). (c) With Vibration + Compensation. We use the EKF-tracked region to estimate and compensate for the sinusoidal motion across the entire scene. The motion-compensated accumulated image recovers sharp structures, and the $y$–$t$ slice aligns with the stable motion-compensated track (magenta), revealing the underlying scene. An inset (bottom left) shows the reference scene displayed in front of the camera with induced motion.
• Figure 2: Schematic of the mass–spring–damper model. Camera setup at two time steps, (a) $t_0$ and (b) $t_1$. The rotation of an off-axis mass $m$ with angular velocity $\omega$ induces planar displacements $\Delta x, \Delta y$ (only the vertical component is illustrated). The estimated oscillation frequency, $\frac{2\pi}{T}$, corresponds to the motion perceived by the camera. Note that this differs from the natural frequency of the off-axis mass $\omega$, due to inertial effects.
• Figure 3: Visual representation of the camera model. The virtual camera $\mathcal{C}$ remains static, while the real camera $\mathcal{C}'$ translates along a circular path centered at $\mathcal{C}$ in the $x$–$y$ image plane of $\mathcal{C}$. We denote $\mathbf{p}$ and $\mathbf{p}'$ as the projected point of $\mathbf{P}_W$ in the virtual and and real image planes respectively. $\hat{A}$ and $A$ refer to the amplitudes of motion in the virtual and real frames respectively. The objective is to remove the resulting oscillatory motion and recover the projection of point $P_W$ onto the image plane of $\mathcal{C}$.
• Figure 4: Schematic representation of VibES. The input event stream is processed by a tracker that extracts trajectories used to estimate the dominant frequency, amplitude, and phase shift of the oscillatory motion. Once the sinusoidal motion is characterized, an Extended Kalman Filter (EKF) is initialized independently for each axis to track the induced motion. The EKF estimates are then used to compensate for the induced motion in the incoming event stream.
• Figure 5: Shannon entropy and NIQE scores of reconstructed frames of binary accumulated event frames, computed over 10 ms time windows on the Logo real-world scene. The shaded range denotes the minimum and maximum values within every 10-frame window. Temporal screenshots of both the VibES and S-EV reconstructions, generated using E2VID Rebecq19cvprRebecq19pami, are shown on the right. The regions where the S-EV method produces low-quality images are those where the scene is completely static or there is slow motion.