人類動作辨識這項技術在電腦視覺及圖形識別等領域一直是重要的研究議題。這項技術可以廣泛應用在如醫學復健、運動動作訓練、電玩遊戲與監控系統等領域裡。在早期的動作辨識相關研究之所以不普及，主要是因為在設備的成本與使用空間需求上不是一般研究者可以負擔得起。但Kinect 體感攝影機的發布改變了這一現象，其低成本的優點使得大量研究者開始投入動作辨識的相關研究領域。
本論文利用 Kinect 裝置搭配由Windows公司推出的Kinect SDK2.0開發套件所提供的25個人體骨架關節點資訊與深度資訊製作出一套人體動作姿勢的辨識系統，用以解決昂貴的傳統深度攝影機所造成的設備成本問題並使用其中的骨架資訊來加強辨識準確率。本論文為了解決複雜的人體動作而產生的自我遮蔽問題和其他錯誤現象所造成的辨識率下降，故我們使用了兩台的Kinect攝影機來做到一個資料來源的互補，這樣即可以大幅降低因可能的遮蔽現象所造成的辨識誤判。之後我們再進行目標物的特徵擷取利用機器學習的方法建立的辨識系統，用以建立我們的資料庫。
關鍵詞：人類動作辨識、Kinect、機器學習、Multiple Views
A Real-Time
Basketball Action Recognition based on Machine Learning Algorithm in Multi-View
Environments
Human action recognition has been an
important research in computer vision and computer graphics. It is widely used
in entertainment, sports, medical applications and surveillance system. The traditional
motion capture equipment is not usually affordable for normal developer. With
the reasonable price of Kinect camera, low-cost human motion recognition
becomes possible.
In this paper, we use multiple Kinect
sensors and Kinect SDK as the tool to build our human action recognition
system. This solves the problem of action recognition equipment costs. Using
multiple Kinect cameras to solve the judging and correction error problems
(such as self-occlusion and image noise...etc.) and using machine learning
method to classified our features, it can make our
recognition result with higher performance.
In our methods, we also have a detection
of basketball to prevent that the subject is without ball, it makes our works
more reasonable. Above of all, this paper have the action recognition rate to
be more than 90% in real-time usage from three of the trained behaviors, i.e.
right-hand dribble, left-hand dribble, and shooting behaviors.
Keyword—action recognition, multiple Kinects, Kinect SDK, machine learning, skeleton.