Abstract

Standard frame-based cameras impose a fundamental tradeoff between motion blur and exposure time that limits their usefulness for fast-moving robots. Event cameras, originally developed at ETH Zurich under the Dynamic Vision Sensor (DVS) name and now produced commercially by Prophesee and Sony, sidestep this tradeoff entirely. Each pixel fires asynchronously when its local log-luminance changes by a threshold, producing a sparse stream of timestamped polarity events rather than dense image frames. The result is microsecond temporal resolution, high dynamic range, and very low data rates during static scenes - properties that make event cameras compelling for agile drone navigation, high-speed manipulation, and adverse-lighting environments.

Event Camera Fundamentals

An event stream is structurally different from a frame stream and cannot be fed directly into conventional convolutional image pipelines. Practical approaches either accumulate events into time-surface representations or event frames (trading away the temporal resolution advantage) or process the sparse asynchronous stream natively using spiking neural networks or graph-based representations. The University of Zurich’s Robotics and Perception Group has released datasets - most notably MVSEC and the Event Camera Dataset - and the rpg_e2vid and ESIM simulation tools that have become standard starting points for event-camera algorithm development.

Navigation at High Speed

The primary use case driving commercial interest is drone racing and agile inspection flight, where frame rates above 200fps are needed to track features between captures but generate prohibitive data volumes with conventional cameras. Event cameras operating at effective rates of 10,000 events per millisecond have enabled feature tracking at drone speeds exceeding 10 m/s through cluttered indoor environments in work from UZH’s Robotics and Perception Group and from the Skydio research team. Odometry pipelines built on event-based contrast maximization - fitting camera motion to maximize the contrast of an event-integrated frame - have reached accuracy competitive with visual-inertial odometry on moderate-speed trajectories and outperform it significantly above 5 m/s.

Integration Challenges

The sensor hardware has matured faster than the software ecosystem. Event cameras from Prophesee’s Metavision SDK and from iniVation’s DV software provide ROS2 drivers and event stream APIs, but calibration tools, synchronization with IMUs, and integration with existing navigation stacks require more manual effort than frame-camera equivalents. Noise is non-trivial: hot pixels fire continuously, and illumination changes from flickering artificial lights generate events that look identical to motion events. Filtering strategies exist, but they add latency. The field is also bifurcated between research groups that treat event cameras as a standalone modality and those building event-frame hybrid sensors (Sony’s IMX636 stacks both on one die), with the hybrid approach likely to dominate production deployments.

Outlook

The neuromorphic processing angle remains compelling on paper but practically distant. Matching event camera data to Intel’s Loihi 2 or BrainScaleS spiking processors has been demonstrated in lab settings, but the programming models are not ready for rapid algorithm development. In the near term, event cameras will be deployed alongside conventional GPUs with specialized pre-processing that converts the event stream into dense tensor representations, accepting the latency cost of that conversion in exchange for the sensor’s dynamic range and motion-blur immunity.