Dynamic Threshold Spatial-Temporal Filter on FPGAs for Event-Based Vision Sensors

Event-based vision sensors are high-speed, wide dynamic range image sensors with potential applications in domains such as robotics and visual navigation. However, these sensors are sensitive to noise, particularly under low-light conditions, degrading the data quality. Therefore, developing a filter capable of detecting and removing noise from different sources with high accuracy is crucial. Moreover, removing near- edge noise in high-density areas is particularly challenging using conventional methods because of their high spatial-temporal correlation with actual events. We propose a dynamic threshold spatial-temporal filter that detects high- or low-density areas and removes noise. Detection was achieved by counting the number of events occurring within a certain period in the area surrounding each event. Applying an appropriate thresh- old for each density significantly enhanced noise processing accuracy, as reflected by the mean square error and peak signal-to-noise ratio metrics. Moreover, we synthesize digital circuits in a field-programmable gate array and demonstrated a notable reduction in processing time compared to that of the central processing unit-based approach, achieving up to 74- fold faster in processing speed. These findings suggest that the proposed filter can significantly enhance real-time event-based vision systems, particularly in environments with varying noise conditions.