A 33 μW 42 GOPS/W 64×64 Pixels Vision Sensor with Dynamic Background Subtraction for Scene Interpretation

摘要

A 64×64 pixel vision sensor performs adaptive background subtraction and event detection at very low power consumption. The chip is based on a VLSI-oriented vision algorithm, implemented at pixel-level, mimicking the basic process of pre-attentive visual perception. Anomalous pixel behaviors are detected and coded into a 2-bit/pixel. Each pixel integrates two programmable Switched-Capacitors Low-Pass Filters and two clocked comparators, which are fundamental blocks for the execution of the vision algorithm. The 45T square pixel has a pitch of 26μm and a fill factor of 12%. The vision, sensor consumes 33μW at 13/ps and 3.3V. This turns into a computing performance of 42 GOPS/W and 4 GO PS/mm~2, which are values aligned with the most; advanced computational vision sensors.