Neural Network Architectures for the Direct Perception of 3D Motion and Depth

跟踪图像序列中的刚性运动的大规模并行方法可以基于李代数形式。这项研究将扩展这种神经网络可视化分析的理论基础，并通过连接机模拟进行实验验证。关键的洞察力是，通过对源自正则李氏分解的霍夫曼轨道进行局部布线，可以实现对吉布森“光学阵列的流动”中的变化的瞬时跟踪。这类似于重复应用Hough变换，但这里将与一种新技术相结合，以利用视觉流的“动量”。这项工作的灵感来自皮茨和麦卡洛克关于从变换群中提取共性的理论。目标是同时跟踪物体的位置和深度，而不需要首先将它们从视场中分割出来。

A massively parallel approach to the tracking of rigid motions in image sequences can be based on a Lie-algebra formalism. This research will extend the theoretical basis of such neural-network visual analysis, with experimental verification via a Connection Machine simulation. The key insight is that instantaneous tracking of variations in Gibson's "flow of the optic array" can be achieved by local wiring of Hoffman orbits derived from a canonical Lie decomposition. This is similar to repeated application of a Hough transform, but will here be combined with a new technique for exploiting "momentum" of the visual flow. Inspiration for this work comes from Pitts and McCulloch's theory on the extraction of universals from transformation groups. The goal is to track objects simultaneously in position and depth, without first segmenting them from the optic field.