DVS Event Vision: Reading Pixel Changes as Temporal Information
Unlike conventional cameras that repeatedly generate full images at fixed intervals, a DVS (Dynamic Vision Sensor) outputs only brightness changes at each pixel as asynchronous events.
Robowink’s DVS Event Vision research builds a unified processing chain from event acquisition, calibration, and denoising to spatiotemporal representation and feature extraction, with the goal of producing stable event data for downstream AI.
Each pixel independently detects brightness changes and generates an event containing position, time, and polarity information when the change exceeds a threshold.
Continuously detects brightness changes at the pixel level.
An ON or OFF event is generated when the change exceeds the threshold.
Each event records both its spatial location and timestamp.
Forms a time-ordered asynchronous event stream instead of image frames.
Raw event streams are converted into different spatiotemporal representations according to the task and used for feature extraction and learning inputs.
Preserves the position, timestamp, and brightness-change polarity of each event in its raw form.
(x, y, t, p)Projects the timestamps of recent events into space to represent the temporal structure of motion.
S(x,y)=exp(-(t-now)/τ)Accumulates events over short time windows to analyze spatial distribution and change intensity.
Σ Event / ΔtDivides the time axis into multiple bins and converts events into a three-dimensional learning input.
x × y × time-binThe value of DVS begins with how the data is generated rather than with a single performance figure from a specific product. Robowink studies how these sensing characteristics can be applied to real industrial data processing.
Each pixel generates events independently without full-frame synchronization.
Represents the timing of motion changes with fine temporal information.
Uses brightness-change-based sensing to capture dynamic information in environments with large illumination differences.
Reduces repeated output from static backgrounds and generates data centered on meaningful changes.
DVS is better viewed as a sensor that complements RGB cameras with temporal-change information rather than as a direct replacement. Understanding the data characteristics of each approach is important when selecting or fusing them for a given application.
The research pipeline on this page focuses on improving the quality and representation of event data itself rather than on SNN inference or control.
Raw event acquisition
Sensor variation calibration
Background and isolated-event removal
Timestamp normalization
Spatiotemporal event grouping
Time Surface / tensor conversion
Quantitative and qualitative validation
We compare and validate algorithms using public event-vision datasets, then evaluate environmental adaptability and reproducibility with event logs acquired in actual industrial environments.
Used to validate algorithms with event-based sensor data from mobile platforms and diverse scenes.
Used as comparative data for event streams in driving environments and for temporal-spatial change analysis.
Used for preprocessing and representation research with data that combines event and image information.
Field event logs acquired in real logistics and manufacturing environments are used to examine the effects of noise, illumination, and motion conditions.
To avoid overlap with the Safety Module in the Business section, this research focuses on foundational technologies for sensor-event processing and event-data representation.
Reduces background activity and isolated events to improve the quality of meaningful event signals.
Groups spatially and temporally related events to extract the structure of dynamic patterns.
Studies task-specific event representations such as Time Surface, Event Frame, and Voxel formats.
Studies foundational technologies for temporal and coordinate alignment between DVS, RGB, and range sensors.
The items below are R&D targets for the event-vision processing chain, not guaranteed performance specifications of a commercial product.
We welcome technical collaboration and PoC opportunities in event-data acquisition, preprocessing, representation, and sensor-fusion research.