在雙眼視覺中利用類神經網路預測3-D 運動軌跡

Abstract

在這篇論文中，我們利用類神經網路為工具，去預測目標物在三度空間中 的運動軌跡。我們先用立體成像(stereo imaging)的方法算出目標物在空 間中的座標，接著採用最小平方法找出片段連續的運動軌跡，將各段軌跡 相對應的參數表示式，當作回傳(back-propagation)類神經網路的訓練對 ，讓網路去學習兩段運動軌跡間的關係。經由訓練完成的網路，我們可以 預測出目標物的運動軌跡，實驗結果顯示我們系統有不錯的預測功能。 In this thesis, we use neural networks to predict motion trajectory of a target in 3-D space. Th stereo imaging method is applied to compute the target 3-D coordinate, then we employ the least square approximation method to decide its piecewise continuous trajectories. The parametric representations corresponding to any two contiguous paths are taken as the training pairs of a back-propagation neural network. Finally the neural network can learn the relation between these two contiguous paths. Through the trained neural network, we can predict the whole target motion trajectoryif the first three target positions are given. Experimental results are provided to demonstrate the proposedscheme with good performance.