Smart Vision Sensors: Analog VLSI Systems for Integrated Image Acquisition and Early Vision Processing

本研究的主要目标是设计和测试用于实时机器视觉应用的模拟VLSI系统。在实时机器视觉中，需要获取、管理和处理的大量图像数据导致成像仪、内存和处理器之间的通信瓶颈，以及非常高的计算需求。目标是确定如何利用模拟VLSI的优点-高速，低功耗和小面积-以及其缺点-有限的精度，不灵活性和缺乏存储容量-可以最小化。这项工作主要集中在早期视觉任务，即发生在动物或机器信号流路径早期的任务。提出的设计包括用于相机速度估计、立体深度、图像分割和平滑、表面重建、大范围亮度自适应成像以及用于晶圆制造的自动基准标记对齐的芯片。麻省理工学院早期的设计表明，典型的子系统在物理上非常小，将以每秒数百到数千帧的速度执行一个或多个计算密集型图像处理任务，仅使用数十到数百毫瓦。

Wyatt The primary goal of this research is to design and test analog VLSI systems for applications of real-time machine vision. In real-time machine vision the sheer volume of image data to be acquired, managed and processed leads to communications bottlenecks between imagers, memory and processors, and to very high computational demands. The goal is to determine how the advantages of analog VLSI--high speed, low power and small area-- can be exploited and its disadvantages--limited accuracy, inflexibility and lack of storage capacity--can be minimized. The work is concentrated on early vision tasks, i.e., tasks that occur early in the signal flow path of animal or machine. Proposed designs include chips for camera velocity estimation, depth from stereo, image segmentation and smoothing, surface reconstruction, wide-range brightness adaptive imaging, and automatic fiducial mark alignment for wafer fabrication. MIT's earlier designs have shown that the typical subsystem is physically very small and will perform one or more computationally intensive image-processing tasks at hundreds-to- thousands of frames per second using only tens-to-hundreds of milliwatts.