基於星狀骨骼特徵之人型姿態辨識系統 設計應用於居家復健

Abstract

近年來，為了因應病患之居家看護及復健需求，在醫療領域中已開發出各種復健輔助器材，能主動監控患者動作之人型姿態辨識系統也逐漸受到重視。本篇論文所提之人型姿態辨識系統除了可幫助患者自行進行復健療程外，同時也可監控患者安全。整個系統由人型影像偵測、特徵擷取及姿態辨識三個部分所組成，在人型影像偵測中，系統使用背景相減、邊緣偵測及物件聯通三個步驟找出人型輪廓影像，接著擷取人型輪廓之星狀骨骼特徵，作為類神經網路分類器的輸入，再利用倒傳遞演算法則來訓練此類神經網路分類器，進而辨識出人型姿態。根據實驗結果，本系統能在0.2秒內完成人型姿態辨識，準確率高達92.39%，可做為即時辨識系統之用。

In recent years, many kinds of recovery auxiliary equipments have been developed in medical field to fulfill the patients’ needs for rehabilitation and security. Human pose recognition system which could automatically supervise patients’ motion has received increasing attention. This thesis proposes a human pose recognition system to help those patients follow the rehabilitation program and supervise their safety. The system is composed of three parts, including human image detection, feature extraction and human pose recognition. For human image detection, it is implemented by background subtraction, edge detection and connected component labeling (CCL). The feature extraction is then processed by the star skeleton method, and the human pose recognition is fulfilled by the neural network classifiers designed by the back propagation algorithm. From the experimental results, the proposed system is able to recognize a human pose in 0.2 second with accuracy 92.39%, which could be applied in real time system.