Study On Event-based CMOS Bionic Vision Sensor

The rapid development of smart society has made the researches of image sensing technique focus on the ability of capturing dynamic scenes.To achieve higher temporal resolution,the frame rate of mainstream CMOS image sensor rises continuously,which leads to the amount of output data grows explosively.However,only a small part of them can represent the dynamic information,while the other is redundancy since it only contains the static backgrounds.The transmission and processing of the redundant data will cause the waste of bandwidth,storage and power consumption.Designing CMOS image sensor with event-based bionic method can solve these problems,which makes it significant to study on event-based CMOS bionic vision sensor for the application of high temporal resolution.In this theis,we analyze the existing CMOS image sensor techniques upon the study of principles of biological vision system and event-based control system,and then determine the architecture of our event-based bionic vision sensor.We also analyze the nonideal factors in the pixel circuit and the error of temporal contrast threshold and time information caused by them.The related models are built after the derivation,and the analysis on the relationship between sensing quality degeneration and the nonidealities are given based on a limited bandwidth Gaussian white noise input.From the analysis,the methods that improve the imaging quality of event-based bionic vision sensor can be obtained.We propose a stacked-subthreshold-MOSFET event-based bionic pixel circuit,which is able to improve the temporal contrast sensitivity,and finally design and test an event-based CMOS bionic vision sensor.Besides,we propose an event-driven sliding method to reconstruct equivalent frames from event stream upon the studies of fixed sliding method and most recent event method.In the experiments,the proposed method shows alleviation of edge missing and trailing compared to the traditional methods.The major innovations in this thesis are as follows:1.We propose a stacked-subthreshold-MOSFET design for the event-based pixel circuit which is sensitive to temporal contrast.Under a typical CMOS process with?_n=0.8,this design can achieve photocurrent-to-voltage conversion gain enlarged by2.25x,and thus improve the temporal contrast sensitivity of the pixel by 50%with the same number of MOSFET and the same capacitance ratio in the pixel.The experimental results are 2.2x and at least 40%,respectively.2.We propose a new type pixel noise,input-dependent noise in the pixel,and establish its mathematical model.This model quantifies the relationship between the value of the noise and the character of input signal and system parameters.With limited bandwidth Gaussian noise input,optimizing the pixel design and application based on the model can improve the sensing quality of event-based bionic vison sensor by at least 20d B.3.We propose an event-driven sliding algorithm to reconstruct equivalent frames from event stream.The algorithm determines the length of sliding window on the foundation of event-driven principles,which makes it change automatically according to the characteristic of event stream.The algorithm can thus alleviate the missing or trailing of the edge in the equivalent frame.