在這篇論文中,我們綜合了體積特徵(volumetric feature)還有空間與時間校正(spatio-temporal alignment)來處理動作識別這個問題。我們使用了適性背景混和模型(adaptive background mixture model)來將人體的部份從影片之中擷取出來,將他們正規化之後依據人體的重心把他們放在每張影格中央。接下來,我們使用了動態時間翹曲(Dynamic Time Warping)配合形狀內容(Shape Context)來作影片方面的時間校正。作完空間與時間校正之後,我們從這段包含著動作的影片中擷取體積特徵來描述表演者的動作。其中體積特徵是由二維空間中的物體偵測裡頭所使用的特徵得到靈感的。為了將二元分類法應用在多類別的問題上,我們在AdaBoost上使用了錯誤更正碼來達到多類別的目的。相對於直觀上每個類別學一個分類器的方法,我們在實驗中證明了這種多類別的學習正確率並不會低於前者。在實驗中我們也闡述這種空間與時間的校正相較於沒校正前,可以大幅提昇辨識率。
In this thesis we use volumetric feature combined with spatial and temporal alignment to deal with action recognition problem. We use the adaptive background mixture model to extract the human body out of the image sequence, normalize and align them in the center of the frame according to the centroid of figure. After that we use Dynamic Time Warping to achieve the temporal alignment, by using of a simplified version of Shape Context. Then we apply the volumetric feature inspired by 2D rectangle feature in object detection on static images. To solve the multi-class learning problem, we apply an multi-class approach of Adaboost by using error-correcting code, which is more effective than one-against-all approach. In the experiment, we demonstrate the using of spatial and temporal alignment can avoid the time-scale and space-scale issue thus improve the accuracy rate.