Evaluating noise filtering for event-based asynchronous change detection image sensors

Bio-inspired Address Event Representation change detection image sensors, also known as silicon retinae, have matured to the point where they can be purchased commercially, and are easily operated by laymen. Noise is present in the output of these sensors, and improved noise filtering will enhance performance in many applications. A novel approach is proposed for quantifying the quality of data received from a silicon retina, and quantifying the performance of different noise filtering algorithms. We present a test rig which repetitively records printed test patterns, along with a method for averaging over repeated recordings to estimate the likelihood of an event being signal or noise. The calculated signal and noise probabilities are used to quantitatively compare the performance of 8 different filtering algorithms while varying each filter's parameters. We show how the choice of best filter and parameters varies as a function of the stimulus, particularly the temporal rate of change of intensity for a pixel, especially when the assumption of sharp temporal edges is not valid.