- 學位論文
透過穿戴式魚眼相機的姿勢偵測
Cyclops: Wearable Fisheye Cameras for Capturing Gestures
摘要
此篇論文提出了使用魚眼相機拍到的影像來做姿勢輸入方式，透過魚眼非常廣的視野可以觀察使用者的姿勢同時也可以讓使用者與環境做互動。根據相機的型態(彩色或紅外線相機)所拍攝到的影像可以簡單的透過標準的影像處理(皮膚顏色偵測或設門檻的方式)來取出前景的部分，可以進一步將取出的前景區域累積起來形成具有時間及空間資訊的樣板(Motion History Images)來偵測有動作的姿勢，我們使用了一種機器學習的技術(Randomized Decision Forests)來辨識這些姿勢。基於上述的架構我們實作了兩個概念性證明的裝置，分別透過穿戴於胸前及手指的裝置來偵測全身性和手部的姿勢。第一個我們稱為Cyclops的裝置中，是透過魚眼鏡頭放置胸前這類身體中心的位置來觀察使用者的四肢，這樣的方式有別於過去使用外在相機或者將移動感知器散佈於使用者身上的方式，Cyclops是單件穿戴型裝置並且可以透過視野很大的魚眼影像來偵測使用者全身性的姿勢。而第二個稱作CyclopsRing的裝置是透過把魚眼鏡頭穿戴在手指邊緣來偵測手部的姿勢還能將手掌及手指當作觸碰介面，別且因為魚眼影像帶來的好處使得使用者可以透過手勢與環境做互動。Cyclops和CyclopsRing分別在有20個健身動作及7個手勢的實驗中達到了92%跟84%的辨識率。
並列摘要
This thesis introduces a novel gesture input system via the images captured by a fisheye lens whose extremely wide field-of-view significantly broadens the system’s capability of perceiving the gestures performed by users and further enables the possibility of interacting with the environment. According to the type of camera sensors (e.g. color or infrared), the captured images can be easily processed to extract the foreground regions by standard image processing techniques (e.g. skin color detection or thresholding) to represent the gestures performed by users. And furthermore, we can enable motion gestures by accumulating the foreground images to temporal templates (e.g. motion history images). To recognize the gestures represented by the image, we use the a machine learning technology called randomized decision forests (RDF). Based on the above framework, we present two proof-of-concept implementations in this thesis. We demonstrate that full-body and whole-hand inter- actions are enabled by realizing the proposed concept with a chest-worn and finger-worn device, respectively. In our first implementation, which we call Cyclops, we positioned the fisheye lens at the central location of the user’s body to see the user’s body limbs. Unlike existing body gesture input systems that requires external camera or distributed motion sensors across the user’s body, Cyclops is a single-piece wearable device and can see the user’s whole body postures from the wide field-of-view fisheye image. And in the second implementation, which we call CyclopsRing, the fisheye lens is worn at hand webbings to see user’s hand from central location that enables hand gesture input and is able to turn skin regions on the fingers and palm into touch interface. Benefitting from the fisheye view, CyclopsRing can incorporate real-world elements into hand-based interactions. CyclopsRing is a ring-stye fisheye imaging device enabling whole-hand and context-aware interaction. The experiment consisting of 20 body workout gestures for Cyclops and 7 hand gestures for CyclopsRing reported 92% and 84% recognition rate achieved by Cyclops and CyclopsRing using RDF respectively.