針對移動設備之記憶體高效RGBD視覺里程計設計

Abstract

隨著移動設備的發展，尤其是增強現實（AR）和虛擬現實（VR）的應用，對能在這些設備的資源限制下運行的高效視覺里程計（VO）算法的需求日益增加。本文提出了一種記憶體高效的RGB-D視覺里程計（RGBD VO）解決方案，通過將算法優化與硬體協同設計相結合，實現減少了十倍的記憶體使用量。該方法在準確度與性能之間取得平衡，適用於資源受限平台上的實時應用。 我們介紹了一種記憶體高效的RGBD VO算法，該算法消除了影像金字塔的需求，顯著降低了記憶體消耗。所提出的硬體架構將記憶體使用量優化了32.78%，將晶片面積縮小了12.25%，並將性能提升了1.68倍，相較於現有方法。基於台積電40nm技術的實現，在處理深度圖時，其數據處理量是傳統方法的近三倍。 本文使用廣泛認可的RGB-D基準測試數據集對該方法進行了評估，結果表明，我們的方法在準確度上與傳統方法相當，甚至略優。通過結合基於特徵和基於直接法的優化技術，我們的方法在最小化記憶體使用的同時確保了姿態估算的高精度，儘管記憶體需求有所減少。 這項工作為實時視覺里程計提供了一個高效的解決方案，特別適用於移動設備和AR/VR應用，這些應用中對記憶體和計算能力有較高的要求。該方法不僅滿足實時需求，還顯示出在下一代移動設備和沉浸式技術中的廣泛應用潛力。

The growth of mobile devices, especially for augmented and virtual reality (AR/VR), has increased the need for efficient visual odometry (VO) algorithms that can operate within limited device resources. This thesis presents a memory-efficient RGB-D visual odometry (RGBD VO) solution that combines algorithmic optimization with hardware co-design, achieving a tenfold reduction in memory usage. The method balances accuracy and performance, making it ideal for real-time applications on resource-constrained platforms.

We introduce a memory-efficient RGBD VO algorithm that eliminates the need for image pyramids, significantly reducing memory usage. The proposed hardware architecture optimizes memory by 32.78%, reduces chip area by 12.25%, and improves performance by 1.68x compared to existing methods. Implemented using TSMC 40nm technology, it processes nearly three times the data, especially when handling depth maps, compared to conventional approaches.

The thesis evaluates the method using a widely recognized RGB-D benchmark dataset, showing that our approach achieves a comparable or slightly better accuracy than traditional methods. By combining feature-based and direct-based optimization techniques, our method minimizes memory usage while ensuring high accuracy in pose estimation, despite the reduced memory requirements.

This work provides an efficient solution for real-time visual odometry, especially for mobile and AR/VR applications, where memory and processing power are limited. The approach meets real-time requirements and demonstrates potential for use in next-generation mobile devices and immersive technologies.