以視覺為基礎的人機互動(Human-Machine Interaction，HMI)技術需藉助一系列複雜的影像處理流程，因而需要高速的處理器才足以因應演算法的實作，對於硬體資源有限的嵌入式系統，要實現即時人機互動的影像處理有其困難度。本研究提出一個即時的動態物件偵測嵌入式硬體架構，再將之應用以手勢辨識。這個方法從連續影像適應性地建立背景模型，並使用連通物件技術標定目標物件。在手勢辨識的應用中，目標物件就是運動手勢。 我們採用MIAT嵌入式系統設計方法論，將動態物件偵測演算法設計成嵌入式硬體，以大幅提高系統即時性能。然後將得到的動態物件資訊傳入手勢辨識系統進行分析，透過動態手勢軌跡變化特徵，結合模糊神經網路(Fuzzy Neural Network, FNN)推論系統進行動態手勢辨識，最後取得辨識結果提供人機互動的指令。使用者可以利用本系統擴充自訂手勢指令以增加其應用範疇和客製功能。我們所實現的即時動態物件偵測硬體加速引擎工作時脈可達107.63MHz，估測其效率可達每秒350張640×480影像的效能，相較於軟體系統，本研究使用低成本的硬體即可滿足即時嵌入式系統需求。 Vision-based human-computer interaction (HMI) technology requires a series of complex image processing and a high-speed processor is necessary for implementing those algorithms. For embedded hardware system, it is even harder to achieve a real-time human-computer interaction system which is based on image processing. For these reasons, we proposed a design of real-time hardware motion object detection engine and the application of gesture recognition. First, a background model is established adaptively from continuous images. Then, the target object will be pointed out with connected component. In gesture recognition, the target object is the moving hand gesture. In implementation, the MIAT embedded system design methodology is applied to the hardware motion object detection engine to improve the system performance. Afterwards, the moving object information will be sent to the gesture recognition system. The fuzzy neural network (FNN) is also applied in the gesture recognition system for analyzing the dynamic gesture trajectory features. Finally, the results of gesture recognition are provided for the instructions of human-computer interaction. Beside, in order to increase the scope of application and customization, we designed a user interface for users to expand their own gesture command in the system. The proposed hardware motion object detection engine can work up to 107.63MHz of the system clock, which is equivalent to approximately 350 fps using images with 640 × 480 dpi. Comparing to software systems, our system can easily meet the needs of real-time embedded systems with low-cost hardware.
