可見光視線追蹤器之眼角定位硬體架構

Abstract

本研究針對僅採單一可見光相機的眼動儀，提出一套高效能、可即時運作的完整視線追蹤硬體系統。由於可見光環境下無法直接取得瞳孔或光斑，本系統改以內眼角作為穩定參考點，整合眼角偵測、眼球模型參數求解、幾何關係推算與動態縮放補償機制，建構完整的視線追蹤硬體架構。在既有高精度眼角偵測演算法的基礎上，我們將完整眼動儀系統硬體化，透過平行運算與管線化設計，在有 限現場可程式化邏輯閘陣列資源內實現快速的眼角定位與視線估測，有效解決使 用者頭部前後移動時的追蹤誤差問題。實驗結果顯示，眼角偵測模組在現場可程 式化邏輯閘陣列平台上的處理速度達每秒5,115幀，而完整的眼動儀系統整合後 仍能達到每秒476幀的處理效能，均具備即時運作能力。系統硬體資源使用率為 百分之十三邏輯元件與百分之三十五記憶體，證明本架構能為可見光眼動儀提供 可靠且高效的完整硬體解決方案。This study proposes a high-performance, real-time complete gaze tracking hardwaresystem for eye trackers using only a single visible light camera. Since pupils or light spotscannot be directly obtained in visible light environments, this system uses the inner eye corneras a stable reference point, integrating eye corner detection, eyeball model parameter solving,geometric relationship calculation, and dynamic scaling compensation mechanisms toconstruct a complete gaze tracking hardware architecture. Based on existing high-precisioneye corner detection algorithms, we implemented the complete eye tracker system inhardware through parallel computing and pipeline design, achieving rapid eye cornerlocalization and gaze estimation within limited field-programmable gate array resources,effectively solving tracking errors caused by users' forward and backward head movements. Experimental results show that the eye corner detection module achieves a processing speed of 5,115 frames per second on the field-programmable gate array platform, while the complete integrated eye tracker system still maintains a processing performance of 476 frames per second, both possessing real-time operation capabilities. The system hardware resource utilization is 13% logic elements and 35% memory, proving that this architecture can provide areliable and efficient complete hardware solution for visible light eye trackers.