DVS+SNN Safety Module for Earlier Motion Awareness
A DVS (Event Camera) detects motion changes as asynchronous events, while an SNN (Spiking Neural Network) analyzes approach direction and temporal dynamics to support AMR risk assessment.
Integrated with the existing AMR controller, the module provides risk information for stop, slow-down, and rerouting decisions, targeting dynamic objects such as people, carts, and loads in logistics and manufacturing environments.
The architecture combines event-driven input with spike-based inference instead of frame-centric image processing, enabling dynamic-risk assessment close to the field edge.
It asynchronously receives events from pixels where brightness changes occur and forms motion-focused input data. This stage rapidly captures changes from dynamic objects such as people and carts.
This AI processing stage converts events into spikes and uses temporal change patterns to infer whether an object is approaching, its approach direction, and potentially hazardous approach conditions.
Event processing, SNN inference, and risk calculation are performed at the edge, with Stop / Slow / Detour decision information delivered to the AMR Controller.
DVS events are filtered and approach features are extracted before SNN inference and risk calculation, after which the results are delivered to the AMR controller.
Asynchronous DVS event input
Background and noise event removal
Extraction of candidate motion regions
Approach-direction and change analysis
Dynamic-risk class inference
TTC, speed, and direction fusion
Stop, slow-down, and rerouting integration
Risk information is generated by considering motion density, approach velocity, directional change, and stopping distance together rather than relying only on object presence.
Assesses motion activity using the amount of event change within the ROI.
Calculates approach-velocity features from changes in position and size over time.
Analyzes the relative change between the AMR heading and the object movement direction.
Calibrates risk-assessment criteria using AMR speed and stopping-distance conditions.
Risk-assessment results are mapped to simple control states that can be used by the AMR Controller.
Rather than replacing an AMR’s core autonomous-navigation system, the DVS+SNN Safety Module is designed as an additional perception layer that complements existing sensors and control systems.
Vehicle-state information such as AMR speed, operating mode, and stopping distance can also be used to tune the risk logic for site-specific conditions.
Field validation focuses on data quality, risk perception, responsiveness, system integration, and scenario repeatability in actual AMR environments.
We verify whether a stable Event Stream can be acquired under changes in lighting, background motion, installation angle, and distance.
We evaluate hazardous-situation classification and decision stability based on the temporal approach patterns of workers, carts, and moving objects.
We measure processing latency from Event Input through SNN inference to Risk Output and evaluate whether responsiveness is sufficient for AMR operation.
We verify that Risk / State information is reliably integrated with the existing AMR Controller and can be used by the control logic.
Dynamic situations that may occur in real operation are defined as repeatable test conditions so that result reproducibility can be evaluated.
The figures below are current R&D targets used to quantify the preceding PoC validation items; they are not guaranteed product specifications.
We review the PoC configuration and integration approach for the sensor, Edge AI, SNN inference, and control interface based on the AMR environment and dynamic-obstacle scenarios.
