- 學位論文
基於多層次高效監督下降預測演算法的強健多模板追蹤設計
Design of Robust Multi-Template Tracking Based on a Multi-Layered Efficient Supervised Descent Prediction Algorithm
摘要
平面物體偵測與追蹤是擴增實境(Augmented Reality)系統中重要的基礎技術。本論文提出一高效且強健的多模板追蹤方法，其包含多個隨機蕨平面偵測器及平面追蹤器，進行平行化運算來達成多平面物體的偵測與追蹤應用。在隨機蕨平面偵測器的設計上，因所訓練出來的隨機蕨分類器佔用相當龐大的記憶體用量，大大限制了系統可偵測的平面物體數量。為了克服此問題，本論文透過最佳化半樸素貝葉斯分類器的參數設定方式，大幅降低每個平面偵測器所需要的記憶體用量，進而實現記憶體高效的多平面偵測系統。在平面追蹤演算法設計上，本論文亦提出一高效監督式下降預測(Efficient Supervised Descent Prediction)方法，其透過監督式下降法(Supervised Descent Method)及高效二階最小化(Efficient Second-order Minimization, ESM)，事先對每個平面物體學習一個多層次線性預測器。在線上追蹤時，則使用預測方式進行強健且精確的平面追蹤效果，讓原本需要計算大量海森矩陣的ESM方法加速至即時追蹤的效果，使得平面追蹤可以兼顧強健性及實時性。實驗結果顯示本論文所提出的方法與現有的平面追蹤方法比較上，不但擁有較高的追蹤成功率及較小的追蹤誤差之外，在影像有強烈雜訊影響下，仍然保有強健的追蹤效果。此外，本論文所提出的方法處理一個平面物體所需的時間僅需3.54ms，因此可達到即時多平面物體的追蹤效果，進而強化擴增實境系統的實用性。
並列摘要
Planar object detection and tracking are important foundational techniques in augmented reality systems. In this thesis, an efficient and robust multi-template tracking method that includes multiple random-ferns planar detectors and real-time planar trackers is proposed for parallelization of multi-template detection and tracking. In the design of the random-ferns planar detector, the trained random-ferns classifier usually occupies a large amount of memory, which greatly limits the number of planar objects detectable by the system. To overcome this problem, this thesis optimizes the parameter settings of a semi-naïve Bayesian classifier to greatly reduce memory usage required by each planar detector, which helps to realize a memory-efficient multi-planar detection system. In the design of the planar tracking algorithm, this thesis proposes an efficient supervised descent prediction method, which learns a multi-layered linear predictor for each planar object in advance based on supervised descent method and efficient second-order minimization (ESM). In the online tracking stage, the proposed supervised descent prediction approach is used for robust and accurate planar tracking, which accelerates the ESM method that originally requires large computational costs on Hessian matrix calculation to achieve fast tracking performance, making planar tracking both robust and real-time. Experimental results show that compared with two existing planar tracking methods, the proposed method not only has a higher tracking success rate and smaller tracking error, but also provides a strong tracking robustness against the influence of high-variance noise in the image. In addition, the proposed method takes only 3.54ms to track a planar object, so simultaneously real-time tracking of multiple planar objects can be achieved, thereby enhancing the practicability of augmented reality system.