RIA: VLSI Architectures and Algorithms for Motion Estimation Using Hough Transform

查克拉巴蒂在过去，实时图像处理一直局限于低级别的任务，如过滤；更复杂的中高级任务已留给无法处理这些任务的大计算速度的通用计算机。因此，这些任务的实时计算需要开发特定于任务的算法和体系结构。该项目正在为一项重要且具有代表性的中级任务开发高效的算法和架构：2维和3维对象的运动估计。该方法通过在Hough空间中实现运动估计，减少了运动估计的计算量。具体地说，目标由霍夫空间中的峰值表示，并从这些峰值的移位中提取运动参数。为了便于Hough变换的实时计算，正在开发减小Hough变换参数空间大小的高效算法。为了保持较低的设计成本，这些算法针对的是定制版本的“基本”架构：一个高度流水线的处理器线性阵列，可以访问一组内存块。除了开发一种高效的实时运动估计方案外，本研究还将为其他中高级图像处理任务的算法和体系结构开发提供设计指导。

Chakrabarti In the past, real-time image processing has been restricted to low-level tasks such as filtering; the more complex medium and high-level tasks have been left to general-purpose computers which are unable to handle the large computation rates of these tasks. Real-time computation of these tasks thus necessitates development of task-specific algorithms and architectures. This project is developing efficient algorithms and architectures for an important and representative medium-level task: motion estimation of 2-dimensional and 3-dimensional objects. The approach reduces the number of computations in motion estimation by implementing it in the Hough space. Specifically, the objects are represented by peaks in the Hough space, and motion parameters are extracted from shifts of these peaks. To facilitate real-time computation of Hough transforms, efficient algorithms that reduce the size of the Hough transform parameter space are being developed. To keep the design costs low, these algorithms are targeted to customized versions of a "base" architecture: a highly pipelined linear array of processors, which has access to an array of memory banks. Apart from developing an efficient scheme for real-time motion estimation, this research will also provide design guidelines for algorithm and architecture development of other medium- and high-level image processing tasks.