Feature Space Trajectory Image Processor

考虑了一种新的特征空间轨迹（FST）描述对象或对象部分的三维变形版本，其中不同的变形对象视图是FST上的不同顶点。相邻的顶点由直线连接。不同的对象或对象部件具有不同的fst。对于输入数据（对象或对象部分），最接近输入数据的FST表示对象的类别，FST上最接近的线段表示对象或部分的位姿。神经网络处理器执行所有需要的距离计算。解决了与这种扭曲物体的新表示相关的基本问题。处理输入数据的时间序列及其姿态估计能力将用于主动视觉系统。本文将研究其在小训练集上的良好泛化和性能。它拒绝错误（非对象）输入和学习新输入的能力将被研究。

A new feature space trajectory (FST) description of 3-D distorted versions of an object or object part is considered in which different distorted object views are different vertices on the FST. Adjacent vertices are connected by straight lines. Different objects or object parts have different FSTs. For input data (an object or an object part), the FST closest to the input data denotes the class of the object and the closest line segment on that FST denotes the pose of the object or part. A neural net processor performs all needed distance calculations. Fundamental issues associated with this new representation of distorted objects are addressed. Processing a time sequence of input data and its pose estimation ability will be used in active vision systems. Its good generalization and performance on small training sets will be studied. Its ability to reject false (non object) inputs and to learn new inputs will be studied.