A comparative analysis of dynamic vision sensors using 180 nm CMOS technology

Abstract

The development of dynamic vision sensors (DVS) is regarded as one of the most relevant advances in CMOS camera focal-plane signal processing, because it is based on neural processing. The type of pixel that is used in a DVS mimicks the functionality of a neural pathway known as magno-cellular pathway, which is responsible for part of the communication between the biological retina and the central nervous system. The magno-cellular pathway responds in asynchronous fashion to light intensity temporal variations, and it encodes such variations by means of neural spike sequences. In this work, we designed and compared three DVS architectures: basic DVS, ATIS (asynchronous time-based image sensor)and ADMDVS (asynchronous delta modulation dynamic vision sensor). Among these architectures, only ATIS implements a light intensity encoding system, using time-based pulse-width modulation. In the design process, gm/ID methodology is used as a suitable tool for pixel design. Using different programming languages, several scripts are developed for making the simulation stages automatic. To verify the correct operation of each architecture, we compare electrical simulation results to numerical simulation predictions that were previously obtained using ideal pixel models. We finally conclude that the behavior of each architecture, which was obtained by electrical simulation, approximates rather well the behavior that was predicted using ideal models, which validates the proposed pixel design for all three sensor types. Based on these results, the basic DVS, ATIS, and ADMDVS architectures may be compared.