壓縮感測技術應用於物件辨識與追蹤

Compressive Sensing Reconstruction Methods for Object Recognition and Tracking

蕭一豪(Hsiao, Yi Hao)

指導教授 ： 陳朝欽

國立清華大學/電機資訊學院/資訊工程學系所/博士(2016年)

摘要

壓縮感知(compressive sensing, compressed sensing)是一新穎的信號處理技術，透過重建演算法用遠低於Nyquist-Shannon取樣定理的取樣頻率來完整重建原始信號。近年來，壓縮感知已被廣泛應用於各領域，本論文基於壓縮感知的稀疏表示與重建演算法概念，提出優化方法應用於物件辨識與追蹤，以獲得較佳的精確性。
一種基於壓縮感知的稀疏表示分類法(sparse representation-based classification, SRC)近期被提出來應用於人臉辨識，本論文基於稀疏表示分類法概念，提出一個局部排序最大概率方法(SRC-maximum probability of the partial ranking, SRC-MP)，所提方法應用於人臉與魚類資料庫，實驗結果顯示提出的方法比基於投影方法：如主成分分析(PCA)、線性判別分析(LDA)、2DPCA、2DLDA，與匹配追蹤(Matching pursuit, MP)相關方法：如正交匹配追蹤(Orthogonal matching pursuit, OMP)、壓縮採樣匹配追蹤(Compressive sampling matching pursuit, CoSaMP)、子空間追蹤(subspace pursuit, SP) 與正則正交匹配追蹤(regularized OMP, ROMP)，可達到更佳的辨識精準度。
壓縮追蹤(compressive tracking)是一個基於壓縮感知的高效即時追蹤方法，本論文基於壓縮追蹤與稀疏表示分類法概念，提出一個樣本蒐集與稀疏樣本表示方法(sparse sample collection and representation, SSCR)，在樣本蒐集方法上，透過整合背景相減法與壓縮追蹤來提升樣本蒐集的準確性。在疏稀樣本表示方法上，將預測樣本以正樣本與負樣本的稀疏方式來表示，並透過重建演算法計算每個預測樣本權重，具最大權重係數的預測樣本即為追蹤結果。同時，在重建演算法部分，本論文提出一個權重調整與動態更新正交匹配追蹤(re-weighting and dynamically updating OMP, RwDuOMP)方法來提升重建效能。所提方法被應用於具複雜環境的台灣墾丁真實世界水下影片，實驗結果顯示提出的方法有效增進魚類追蹤準確度。

關鍵字

壓縮感測 ； 物件追蹤 ； 物件辨識

並列摘要

Compressive sensing (Compressed sensing, CS) is a novel sampling technique which adopts reconstruction algorithms to reconstruct original signals from significantly fewer samples than those using the Nyquist-Shannon sampling theorem. Recently, several researches have been conducted to apply the CS framework to various applications. In this thesis, we propose improved reconstruction methods based on CS for object recognition and tracking.
In recent years, a sparse representation-based classification (SRC) method based on CS is presented for robust face recognition. Our first proposed enhancement is adopting a maximum probability of the partial ranking method based on the framework of SRC, called SRC-MP. It computes the maximum probability from the largest γ weighting coefficients for the subjects. The criterion of selection is now based on the maximum probability, instead of the largest weighting coefficients. Experiments are implemented on face and real-world fish databases. Experimental results show that our proposed method is able to achieve higher accuracy than projection-based methods, such as principal component analysis (PCA), linear discriminant analysis (LDA), 2DPCA and 2DLDA, and matching pursuit related algorithms, such as orthogonal matching pursuit (OMP), compressive sampling matching pursuit (CoSaMP), subspace pursuit (SP), and regularized OMP (ROMP).
On the other hand, a real-time compressive tracking (CT) method based on CS is proposed for object tracking. Our proposed enhancement is implemented for a sparse sample collection and representation (SSCR) method, based on CT and SRC concepts, for real-world fish tracking. The SSCR consists of sample collection and sparse sample representation procedures. The sample collection procedure incorporates background subtraction into CT to improve the accuracy of collecting sets of three kinds of samples (positive, negative, and predictive). The sparse sample representation procedure represents each predictive sample as a sparse linear combination of all positive and negative samples. The weights of the predictive samples are computed using our proposed re-weighting and dynamically updating orthogonal matching pursuit (RwDuOMP) method. The RwDuOMP method includes three procedures, picking over samples, re-weighting the picked samples, and dynamically updating negative samples. The predictive sample with the maximum weighting coefficient is regarded as the target object tracking result. We evaluate the SSCR method using several challenging real-world underwater sequences from an uncontrolled open sea in Taiwan. In addition, we compare the RwDuOMP method with OMP, CoSaMP, SP and ROMP methods. Experimental results indicate that our proposed method improves the accuracy of fish tracking.

並列關鍵字

Compressive Sensing ； Object Tracking ； Object Recognition

參考文獻

[Avi2004] S. Avidan, “Support Vector Tracking,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 26, no. 8, pp. 1064-1072, 2004.

[Bar2007] R.G. Baraniuk, “Compressive Sensing,” IEEE Signal Processing Magazine, vol. 24, no. 4, pp. 118-121, 2007.

[Can2005] E. Candes, and J. Romberg, “l1-MAGIC: Recovery of Sparse Signals via Convex Programming,” Techical Report, Caltech, October, 2005.

[Can2006] E. Candes, J. Romberg, and T. Tao, “Stable Signal Recovery from Incomplete and Inaccurate Measurements,” Communications on Pure and Applied Mathematics, vol. 59, no. 8, pp. 1207-1223, 2006.

[Can2008A] E. Candes, and M. Wakin, “An Introduction to Compressive Sampling,” IEEE Signal Processing Magazine, vol. 25, no. 2, pp. 21-30, 2008.