A 128×128 120dB 30mW Asynchronous Vision Sensor that Responds to Relative Intensity Change

摘要

The frame-based architectures of most imagers are natural for making movies and pictures, they have significant drawbacks for machine vision. Short-latency vision problems require high frame rate, producing massive readout (e.g., >1GB/s from 352×288 pixels at 10kPrames/s [1]). Reducing the output to a manageable rate by using region-of-interest readout usually requires complex control strategies. Readout and processing of largely redundant data ultimately limit reductions in computational effort and power consumption. In this paper, a vision sensor is presented whose pixels asynchronously respond to events that represent relative changes in intensity. It operates largely independent of scene illumination, directly encodes object reflectance, and reduces redundancy while preserving precise timing information. Because output bandwidth is automatically dedicated to dynamic pairs of the scene, the sensor is suitable for applications in surveillance and motion analysis. It improves on prior frame-based temporal difference detection imagers (e.g., [2]) by asynchronously responding to temporal contrast rather than absolute illumination, and on prior event-based imagers because they either do not reduce redundancy at all [3], reduce only spatial redundancy [4], have large FPN, slow response, and limited dynamic range [5], or have low contrast sensitivity [6].