智慧型運動系統之設計

Abstract

本論文主要目的是設計智慧型運動描述系統，此系統是根據人腦所認知運動狀態來識別不同的運動事件，其三種運動事件包含"拿起物件"，"放下物件' 和 '丟下物件'可以被此系統描述。運動狀態的認知以及運動事件的識別可以透過類神經網路的學習來達成。在運動事件的識別上，本系統分成兩部分來處理，第一部分是觸發式類神經網路，第二部分是運動的分類器。由於運動事件和以觸發的運動狀態序列有關，藉由觸發式類神經網路可以把運動狀態序列轉成觸發的運動狀態序列.運動的分類器包含前饋式類神經網路和遞回式類神經網路，然而含前饋式類神經網路會受未定運動狀態序列影響，所以在設計上必須將未定運動狀態序列考慮，而遞回式類神經網路在設計上會根據不同觸發狀態條件會給學習暫態樣。最後，這兩種架構都可達到很好的運動事件的識別效能且不受未知運動狀態序列影響。

The main purpose of this thesis is to design of intelligent motion description system. This system can recognize different motion event according to cognitive motion state of human brain. There are three motion event including “Picking object”, “Putting object” and “Dropping object” which can be described by this system. The cognition of motion state and recognition of motion event can be learned by neural network. A motion event analyzer composed of trigger net and motion classifier applied to recognition of motion event. The motion event is closely related to change of motion state; therefore, a trigger net, applied here, used to turn motion state sequence into triggered state sequence then the triggered state sequence can be classified by motion classifier. Motion classifier can be implemented by two types of neural network, feed-forward and recurrent. However, feed-forward classifier is interfered by undefined motion state sequence. Therefore, undefined motion state sequence will be taken into account to prevent misclassification in design of feed-forward classifier. The recurrent classifier learns transient pattern according to different triggered state condition. Finally, the two types of motion classifier can reject the interference caused by undefined motion state sequence and reach better performance for recognition of motion event.