A dynamic vision sensor with direct logarithmic output and full-frame picture-on-demand
Date of Issue2016-12-08
School of Electrical and Electronic Engineering
Centre for Integrated Circuits and Systems
Inspired by the working mechanism of human retina, the dynamic vision sensor (DVS) has become a research hotspot recently because of its certain advantages over the conventional frame-based image sensor. Instead of working in integration mode, the pixel in DVS continuously monitors local light intensity change through a front-end logarithmic photo detector and reports an event if certain threshold is reached. The output of DVS is not frames, but a stream of asynchronous address-events. The speed of the sensor is not limited by any traditional concept such as exposure time and frame rate. It can detect fast motion which is traditionally captured by expensive, high-speed cameras running at thousands of frames per second, but with significantly reduced amount of data. The early DVSs only output binary address-events. In other words, only the row and column address of the detected events are reported by the sensor. The absence of absolute light intensity information, however, severely restricts their applications in lots of image processing algorithms, such as object recognition, classification and so on. This leads to the subsequent trend to integrate DVS with the “snapshot” function. Two pioneering works combine DVS with extra in-pixel light intensity measurement circuits based on pulse width modulation (PWM) and active pixel sensor (APS) scheme respectively. Unfortunately, both designs rely on the time-consuming integration method to achieve intensity information and thus, results in undesired discordance between the very fast address-events output and the postponed intensity readout, which neutralizes the potential advantages of the event-based system. In this thesis, a new DVS which directly takes the voltage of the in-pixel logarithmic photo detector as its intensity information is presented. This voltage responds to incident light continuously and therefore no additional exposure time is needed. Combined with a column-parallel area-saving SAR ADC array, the sensor is capable of reporting the fast address-events as well as their corresponding intensity information simultaneously. Wide intra-scene dynamic range is also achieved due to the logarithmic compression of light intensity. On the other hand, an external control signal is also added which can force all the pixels to produce request signals and in turn a full-array on-demand reference picture can be obtained. This also enables the xiv sensor to output continuous frames, which makes it compatible with mainstream image processing algorithms. By means of some improvements in circuit design at pixel, column and system level, the image quality of the sensor is ensured with favourable signal-to-noise ratio (SNR) and fixed-pattern-noise (FPN). A 160×192 prototype has been implemented using AMS 0.35 𝜇�� 2P4M CMOS technology. The pixel occupies 30×30 𝜇��� with 14.5 % fill factor. The total chip area is 7.3×7.4 𝑚��� and the average power consumption is about 85 𝑚�� with 40 𝑀��� clock.