- 學位論文
應用FPGA於即時手勢辨識系統之設計與實作
Design and Implementation of FPGA-based Real-Time Hand Gesture Recognition Systems
摘要
本文旨在應用硬體實作技術，建立一套即時手勢辨識系統，作為非接觸式人機互動之操作介面。以往在手勢辨識系統之研究，大多以個人電腦為運算平台，開發手勢辨識演算法。這些方法雖然可以藉由大量訓練樣本與複雜計算流程，提昇辨識準確率，但無法滿足即時辨識應用之需求。本研究提出結合區塊光流梯度與膚色偵測之方法，簡化手部移動偵測之計算程序，發展出適用於硬體實作且不受手形變化影響之動態位移偵測技術，同時整合卡爾曼濾波器與有限狀態機於辨識系統中，使本文所提出之系統能有效達成準確且快速之辨識效果。本研究以軟硬體共同設計進行系統架構規劃，實作於Altera Cyclone II EP2C70開發平台，作為本文所提方法之驗證。在僅耗費25%邏輯單元之系統實作下，即可達成解析度640×480像素，每秒30幅影像之動態手勢辨識速度。在辨識準確率方面，以四種不同場景進行六種手勢測試，平均辨識率於室內環境下為100%，室外環境下為95%。實驗結果顯示，本文所提方法在低複雜之硬體架構下，仍可維持高準確之辨識率。最後，本研究將完成之系統實際應用於電視機之控制，結果亦顯示本文所發展之手勢辨識系統能整合於一般市售家電，提供操作簡便與低成本之人機互動功能。
並列摘要
This study applied hardware implementation techniques to develop a real-time gesture recognition system to act as a non-contact human-computer interaction operation interface. Most previous hand gesture recognition related studies used personal computers as computing platforms to develop hand gesture recognition algorithms. Although these methods may improve recognition accuracy when using a large number of training samples and a complex calculation process, they are insufficient for instant recognition application demands. We used the block gradient-based optical flow method and the skin color detection method to simplify the calculation procedures for hand motion detection to develop a dynamic motion detection technology that is suitable for hardware application and is not influenced by hand shape changes. In addition, we integrated the Kalman filter and a finite state machine into the recognition system to effectively enable our system to achieve accurate and fast identification. We co-designed the hardware and software to conduct system architecture planning and implemented the system in the Altera Cyclone II EP2C70 development platform to verify our proposed method. Under a system implementation that only consumed 25% of the logic elements, we attained a resolution of 640×480 pixels and a dynamic hand gesture recognition speed of 30 frames per second. Regarding recognition accuracy rates, we conducted six gesture tests in four different scenarios. The average recognition rates were 100% for an indoor environment and 95% for an outdoor environment. The experimental results indicated that the proposed method can maintain high accuracy recognition rates despite using low-complexity hardware architecture. Finally, we applied the completed system to television controls, and the results indicated that our hand gesture recognition system can be integrated into general household appliances to provide easy to operate and low cost human-machine interaction capabilities.