人體動作辨識之推論與取樣頻率研究

Abstract

人體動作辨識系統在電腦視覺領域一直是很熱門的研究與應用目標。在居家監控系統中最常見的方式是，使用固定式的攝影機，對室內的人物進行追蹤與動作辨識。為了達到即時監控之目標，處理的演算法必須快速，而且又必須能夠有效的分析影像。 在本論文中，動作辨識的目標是人體，為了更正確的擷取出人體部份，我們同時使用灰階域與HSV色彩空間，建立兩個背景模型，提升消除影像中陰影部分之影響，使得前後景之分離結果能夠更完整。取得即時影像，擷取出的前景部份，經過特徵空間轉換與標準空間轉換後，累積三張動作影像後，藉由預先學習而建立之模糊法則與時序動作姿態比對，完成人體動作之辨識。 研究對於較短周期的動作其取樣頻率改變是否獲得更多資訊，更多的訊息可以使人體動作辨識更加的準確，並且對判斷相同動作的規則，取其最大或者前三大、前五大、前七大和前九大相似度的動作法則平均值，藉由更多規則決定目前輸入的影像與判別動作之間的相似度，確能更加準確判斷人體動作。

Human activity recognition system is now a very popular subject for research and application. Using a fixed camera to track a person and recognize his (her) activity is widely seen in home surveillance. For real-time surveillance, the embedded algorithms must be efficient and fast to meet the real-time constraint. In the thesis, we build two background models, one is grayscale another is HSV color space that extract the human region correctly, and we also reduce the shadowing effect. For better efficiency, the binary image is transformed to a new space by eigenspace and canonical space transformation. After that, we gathered three consecutive down-sampled images to recognize the human actions by fuzzy rules. We utilize different down-sampling rate for short-period action to obtain more information which is useful for the human action recognition. Furthermore, we investigate to the average value of maximal top-3, top-5, top-7 and top-9 firing strength of rules with the same action to recognize the human action. Using more rules to determine the similarity between the inputs and rules that can be more accurately determine human action.