Recent advances on low-rank and sparse decomposition for moving object detection
6 likes2,297 views
The document discusses recent advances in low-rank and sparse decomposition techniques for moving object detection, focusing on various matrix and tensor-based approaches. It covers background subtraction methodologies, including traditional and advanced matrix factorization methods such as Robust Principal Component Analysis (RPCA) and tensor decomposition, highlighting their applications in intelligent video surveillance systems. Also presented are various associated libraries and algorithms for matrix completion and object detection.
Recent advances on low-rank and sparse decomposition for moving object detection
- 1. Recent advances on low-rank and sparse decomposition for moving object detection Matrix and Tensor-based approaches Andrews Cordolino Sobral Ph.D. Student, Computer Vision L3i / MIA, Université de La Rochelle http://andrewssobral.wix.com/home Atelier : Enjeux dans la détection d’objets mobiles par soustraction de fond
- 2. Summary ● Context – Understanding an Intelligent Video Surveillance Framework – Introduction to Background Subtraction ● Decomposition into Additive Matrices – Case 1: Low-rank Approximation and Matrix Completion – Case 2: Robust Principal Component Analysis (RPCA) – Case 3: Stable decomposition ● Constrained RPCA ● Introduction to Tensors – Tensor Decomposition ● Tucker/HoSVD ● CANDECOMP-PARAFAC (CP) ● Applications to Background Subtraction
- 3. Behind the Scenes of an Intelligent Video Surveillance Framework ! Video Content Analysis (VCA) Behavior Analysis Image acquisition and preprocessing Object Detection Object Tracking event location Intrusion detection Collision prevention Target detection and tracking Anomaly detection Target behavior analysis Traffic data collection and analysis activity report Understanding an Intelligent Video Surveillance Framework supervisor ! Example of automatic incident detection our focus
- 4. Introduction to Background Subtraction Initialize Background Model frame model Foreground Detection Background Model Maintenance
- 5. Background Subtraction Methods Traditional methods: • Basic methods, mean and variance over time • Fuzzy based methods • Statistical methods • Non-parametric methods • Neural and neuro-fuzzy methods Matrix and Tensor Factorization methods: • Eigenspace-based methods (PCA / SVD) • RPCA, LRR, NMF, MC, ST, etc. • Tensor Decomposition, NTF, etc. BGSLibrary (C++) https://github.com/andrewssobral/bgslibrary A large number of algorithms have been proposed for background subtraction over the last few years: LRSLibrary (MATLAB) https://github.com/andrewssobral/lrslibrary our focus Andrews Sobral and Antoine Vacavant. A comprehensive review of background subtraction algorithms evaluated with synthetic and real videos. Computer Vision and Image Understanding (CVIU), 2014. Bouwmans, Thierry; Sobral, Andrews; Javed, Sajid; Ki Jung, Soon; Zahzah, El-Hadi. "Decomposition into Low-rank plus Additive Matrices for Background/Foreground Separation: A Review for a Comparative Evaluation with a Large-Scale Dataset". Submitted to Computer Science Review, 2015.
- 6. Background Subtraction Methods Traditional methods: • Basic methods, mean and variance over time • Fuzzy based methods • Statistical methods • Non-parametric methods • Neural and neuro-fuzzy methods Matrix and Tensor Factorization methods: • Eigenspace-based methods (PCA / SVD) • RPCA, LRR, NMF, MC, ST, etc. • Tensor Decomposition, NTF, etc. BGSLibrary (C++) https://github.com/andrewssobral/bgslibrary A large number of algorithms have been proposed for background subtraction over the last few years: LRSLibrary (MATLAB) https://github.com/andrewssobral/lrslibrary our focus Andrews Sobral and Antoine Vacavant. A comprehensive review of background subtraction algorithms evaluated with synthetic and real videos. Computer Vision and Image Understanding (CVIU), 2014. Bouwmans, Thierry; Sobral, Andrews; Javed, Sajid; Ki Jung, Soon; Zahzah, El-Hadi. "Decomposition into Low-rank plus Additive Matrices for Background/Foreground Separation: A Review for a Comparative Evaluation with a Large-Scale Dataset". Submitted to Computer Science Review, 2015. Glossary of terms: PCA Principal Component Analysis SVD Singular Value Decomposition LRA Low-rank Approximation MC Matrix Completion NMF Non-negative Matrix Factorization RPCA Robust Principal Component Analysis LRR Low-rank Recovery RNMF Robust NMF ST Subspace Tracking Stable RPCA Stable version of RPCA TTD Three-Term Decomposition TD Tensor Decomposition NTF Non-negative Tensor Factorization
- 7. Decomposition into Additive Matrices ● The decomposition is represented in a general formulation: ● where K usually is equal to 1, 2, or 3. For K = 3, M1 … M3 are commonly defined by: ● The characteristics of the matrices MK are as follows: – The first matrix M1 = L is the low-rank component. – The second matrix M2 = S is the sparse component. – The third matrix M3 = E is generally the noise component. ● When K = 1, the matrix A ≈ L and S (implicit) can be given by S = A – L. e.g.: LRA, MC, NMF, ... ● When K = 2, A = L + S. This decomposition is called explicit. e.g.: RPCA, LRR, RNMF, … ● When K = 3, A = L + S + E. This decomposition is called stable. e.g.: Stable RPCA / Stable PCP. Bouwmans, Thierry; Sobral, Andrews; Javed, Sajid; Ki Jung, Soon; Zahzah, El-Hadi. "Decomposition into Low-rank plus Additive Matrices for Background/Foreground Separation: A Review for a Comparative Evaluation with a Large-Scale Dataset". Submitted to Computer Science Review, 2015.
- 8. Decomposition into Additive Matrices ● The decomposition is represented in a general formulation: ● where K usually is equal to 1, 2, or 3. For K = 3, M1 … M3 are commonly defined by: ● The characteristics of the matrices MK are as follows: – The first matrix M1 = L is the low-rank component. – The second matrix M2 = S is the sparse component. – The third matrix M3 = E is generally the noise component. ● When K = 1, the matrix A ≈ L and S (implicit) can be given by S = A – L. e.g.: LRA, MC, NMF, ... ● When K = 2, A = L + S. This decomposition is called explicit. e.g.: RPCA, LRR, RNMF, … ● When K = 3, A = L + S + E. This decomposition is called stable. e.g.: Stable RPCA / Stable PCP. Bouwmans, Thierry; Sobral, Andrews; Javed, Sajid; Ki Jung, Soon; Zahzah, El-Hadi. "Decomposition into Low-rank plus Additive Matrices for Background/Foreground Separation: A Review for a Comparative Evaluation with a Large-Scale Dataset". Submitted to Computer Science Review, 2015.
- 9. Low Rank Approximation ▪ Low-rank approximation (LRA) is a minimization problem, in which the cost function measures the fit between a given matrix (the data) and an approximating matrix (the optimization variable), subject to a constraint that the approximating matrix has reduced rank.
- 10. Low Rank Approximation ▪ Low-rank approximation (LRA) is a minimization problem, in which the cost function measures the fit between a given matrix (the data) and an approximating matrix (the optimization variable), subject to a constraint that the approximating matrix has reduced rank. !!! Singular Value Decomposition !!!
- 11. Singular Value Decomposition ▪ Formally, the singular value decomposition of an m×n real or complex matrix A is a factorization of the form: ▪ where U is a m×m real or complex unitary matrix, D is an m×n rectangular diagonal matrix with non-negative real numbers on the diagonal, and VT (the transpose of V if V is real) is an n×n real or complex unitary matrix. The diagonal entries D are known as the singular values of A. ▪ The m columns of U and the n columns of V are called the left-singular vectors and right-singular vectors of A, respectively. generalization of eigenvalue decomposition
- 12. Best rank r Approximation
- 14. What about LRA for corrupted entries?
- 15. Introduction to Matrix Completion (MC) ▪ Matrix Completion (MC) can be formulated as the problem of o recover a low rank matrix L from the partial observations of its entries (represented by A): L Underlying low-rank matrix A Matrix of partial observations http://perception.csl.illinois.edu/matrix-rank/home.html
- 16. Introduction to Matrix Completion (MC) ▪ Matrix Completion (MC) can be formulated as the problem of o recover a low rank matrix L from the partial observations of its entries (represented by A): L Underlying low-rank matrix A Matrix of partial observations http://perception.csl.illinois.edu/matrix-rank/home.html
- 17. Demo: Matrix Completion http://cvxr.com/tfocs/demos/matrixcompletion/ Matrix completion via TFOCS
- 19. MC Algorithms ● LRSLibrary: – MC: Matrix Completion (14) ● FPC: Fixed point and Bregman iterative methods for matrix rank minimization (Ma et al. 2008) ● GROUSE: Grassmannian Rank-One Update Subspace Estimation (Balzano et al. 2010) ● IALM-MC: Inexact ALM for Matrix Completion (Lin et al. 2009) ● LMaFit: Low-Rank Matrix Fitting (Wen et al. 2012) ● LRGeomCG: Low-rank matrix completion by Riemannian optimization (Bart Vandereycken, 2013) ● MC_logdet: Top-N Recommender System via Matrix Completion (Kang et al. 2016) ● MC-NMF: Nonnegative Matrix Completion (Xu et al. 2011) ● OP-RPCA: Robust PCA via Outlier Pursuit (Xu et al. 2012) ● OptSpace: Matrix Completion from Noisy Entries (Keshavan et al. 2009) ● OR1MP: Orthogonal rank-one matrix pursuit for low rank matrix completion (Wang et al. 2015) ● RPCA-GD: Robust PCA via Gradient Descent (Yi et al. 2016) ● ScGrassMC: Scaled Gradients on Grassmann Manifolds for Matrix Completion (Ngo and Saad, 2012) ● SVP: Guaranteed Rank Minimization via Singular Value Projection (Meka et al. 2009) ● SVT: A singular value thresholding algorithm for matrix completion (Cai et al. 2008) https://github.com/andrewssobral/lrslibrary
- 20. MC Algorithms ● LRSLibrary: – MC: Matrix Completion (14) ● FPC: Fixed point and Bregman iterative methods for matrix rank minimization (Ma et al. 2008) ● GROUSE: Grassmannian Rank-One Update Subspace Estimation (Balzano et al. 2010) ● IALM-MC: Inexact ALM for Matrix Completion (Lin et al. 2009) ● LMaFit: Low-Rank Matrix Fitting (Wen et al. 2012) ● LRGeomCG: Low-rank matrix completion by Riemannian optimization (Bart Vandereycken, 2013) ● MC_logdet: Top-N Recommender System via Matrix Completion (Kang et al. 2016) ● MC-NMF: Nonnegative Matrix Completion (Xu et al. 2011) ● OP-RPCA: Robust PCA via Outlier Pursuit (Xu et al. 2012) ● OptSpace: Matrix Completion from Noisy Entries (Keshavan et al. 2009) ● OR1MP: Orthogonal rank-one matrix pursuit for low rank matrix completion (Wang et al. 2015) ● RPCA-GD: Robust PCA via Gradient Descent (Yi et al. 2016) ● ScGrassMC: Scaled Gradients on Grassmann Manifolds for Matrix Completion (Ngo and Saad, 2012) ● SVP: Guaranteed Rank Minimization via Singular Value Projection (Meka et al. 2009) ● SVT: A singular value thresholding algorithm for matrix completion (Cai et al. 2008) https://github.com/andrewssobral/lrslibrary
- 21. Demo: LRSLibrary for MC https://github.com/andrewssobral/lrslibrary https://github.com/andrewssobral/lrslibrary/blob/master/algorithms/mc/GROUSE/run_alg.m
- 23. Decomposition into Additive Matrices ● The decomposition is represented in a general formulation: ● where K usually is equal to 1, 2, or 3. For K = 3, M1 … M3 are commonly defined by: ● The characteristics of the matrices MK are as follows: – The first matrix M1 = L is the low-rank component. – The second matrix M2 = S is the sparse component. – The third matrix M3 = E is generally the noise component. ● When K = 1, the matrix A ≈ L and S (implicit) can be given by S = A – L. e.g.: LRA, MC, NMF, ... ● When K = 2, A = L + S. This decomposition is called explicit. e.g.: RPCA, LRR, RNMF, … ● When K = 3, A = L + S + E. This decomposition is called stable. e.g.: Stable RPCA / Stable PCP. Bouwmans, Thierry; Sobral, Andrews; Javed, Sajid; Ki Jung, Soon; Zahzah, El-Hadi. "Decomposition into Low-rank plus Additive Matrices for Background/Foreground Separation: A Review for a Comparative Evaluation with a Large-Scale Dataset". Submitted to Computer Science Review, 2015.
- 24. Robust Principal Component Analysis (RPCA) ▪ RPCA can be formulated as the problem of decomposing a data matrix A into two components L and S, where A is the sum of a low-rank matrix L and a sparse matrix S: Sparse error matrix SL Underlying low-rank matrix A Matrix of corrupted observations
- 25. Robust Principal Component Analysis (RPCA) Video Low-rank Sparse Foreground Background model Moving objects Classification ▪ Candès et al. (2009) show that L and S can be recovered by solving a convex optimization problem, named as Principal Component Pursuit (PCP):
- 26. Solving PCP One effective way to solve PCP for the case of large matrices is to use a standard augmented Lagrangian multiplier method (ALM) (Bertsekas, 1982). and then minimizing it iteratively by setting where: More information: (Qiu and Vaswani, 2011), (Pope et al. 2011), (Rodríguez and Wohlberg, 2013)
- 27. RPCA solvers For more information see: (Lin et al., 2010) http://perception.csl.illinois.edu/matrix-rank/sample_code.html
- 28. What about RPCA for very dynamic background? http://changedetection.net/ http://www.svcl.ucsd.edu/projects/background_subtraction/demo.htm
- 29. Decomposition into Additive Matrices ● The decomposition is represented in a general formulation: ● where K usually is equal to 1, 2, or 3. For K = 3, M1 … M3 are commonly defined by: ● The characteristics of the matrices MK are as follows: – The first matrix M1 = L is the low-rank component. – The second matrix M2 = S is the sparse component. – The third matrix M3 = E is generally the noise component. ● When K = 1, the matrix A ≈ L and S (implicit) can be given by S = A – L. e.g.: LRA, MC, NMF, ... ● When K = 2, A = L + S. This decomposition is called explicit. e.g.: RPCA, LRR, RNMF, … ● When K = 3, A = L + S + E. This decomposition is called stable. e.g.: Stable RPCA / Stable PCP. Bouwmans, Thierry; Sobral, Andrews; Javed, Sajid; Ki Jung, Soon; Zahzah, El-Hadi. "Decomposition into Low-rank plus Additive Matrices for Background/Foreground Separation: A Review for a Comparative Evaluation with a Large-Scale Dataset". Submitted to Computer Science Review, 2015.
- 30. Stable PCP ▪ The PCP is limited, the low-rank component needs to be exactly low-rank and the sparse component needs to be exactly sparse, but in real applications the observations are often corrupted by noise. ▪ Zhou et al. (2010) proposed a stable version of PCP, named Stable PCP (SPCP), adding a third component that guarantee stable and accurate recovery in the presence of entry-wise noise. The observation matrix A is represented as A = L + S + E, where E is a noise term.
- 31. Constrained RPCA (example 1) ▪ Some authors added an additional constraint to improve the background/foreground separation: – Oreifej et al. (2013) use a turbulance model that quantify the scene’s motion in terms of the motion of the particles which are driven by dense optical flow. http://www.cs.ucf.edu/~oreifej/papers/3-Way-Decomposition.pdf
- 32. Constrained RPCA (example 2) ▪ Yang et al. (2015) propose a robust motion-assisted matrix restoration (RMAMR) where a dense motion field is estimated for each frame by dense optical flow, and mapped into a weighting matrix which indicates the likelihood that each pixel belongs to the background. http://projects.medialab-tju.org/bf_separation/
- 33. Double-constrained RPCA? ▪ Sobral et al. (2015) propose a double-constrained Robust Principal Component Analysis (RPCA), named SCM-RPCA (Shape and Confidence Map-based RPCA), is proposed to improve the object foreground detection in maritime scenes. It combine some ideas of Oreifej et al. (2013) and Yang et al. (2015). – The weighting matrix proposed by Yang et al. (2015) can be used as a shape constraint (or region constraint), while the confidence map proposed by Oreifej et al. (2013) reinforces the pixels belonging from the moving objects. ▪ The original 3WD was modified adding the shape constraint as has been done in the RMAMR. We chose to modify the 3WD instead of RMAMR due its capacity to deal more robustly with the multimodality of the background. https://sites.google.com/site/scmrpca/
- 34. Solving the SCM-RPCA Is important to note that the double constraints (confidence map and shape) can be built from two different types of source (i.e. from spatial, temporal, or spatio-temporal information), but in this work we focus only on spatial saliency maps.
- 35. SCM-RPCA - Visual results on UCSD data set From left to right: (a) input frame, (b) saliency map generated by BMS, (c) ground truth, (d) proposed approach, (e) 3WD, and (f) RMAMR. Dataset: http://www.svcl.ucsd.edu/projects/background_subtraction/ucsdbgsub_dataset.htm
- 36. SCM-RPCA - Visual results on MarDT data set Is important to note that in the UCSD scenes we have used the original spatial saliency map provided by BMS, while for the MarDT scenes we have subtracted its temporal median due to the high saliency from the buildings around the river. Dataset: http://www.dis.uniroma1.it/~labrococo/MAR/index.htm
- 38. What about multidimensional data?
- 40. Introduction to tensors ● Tensors are simply mathematical objects that can be used to describe physical properties. In fact tensors are merely a generalization of scalars, vectors and matrices; a scalar is a zero rank tensor, a vector is a first rank tensor and a matrix is the second rank tensor.
- 41. Introduction to tensors ● Subarrays, tubes and slices of a 3rd order tensor.
- 42. Introduction to tensors ● Matricization and unfolding a 3rd order tensor.
- 43. Introduction to tensors ● Horizontal, vertical and frontal slices from a 3rd order tensor. 1 10 19 28 37 46 51 1 9 17 25 33 41 48 1 9 17 25 33 41 48 kj i 1 10 19 28 37 46 51 1 9 17 25 33 41 48 1 9 17 25 33 41 48 kj i 1 10 19 28 37 46 51 1 9 17 25 33 41 48 1 9 17 25 33 41 48 kj i 1 10 19 28 37 46 51 1 9 17 25 33 41 48 1 9 17 25 33 41 48 kj i 1 10 19 28 37 46 51 1 9 17 25 33 41 48 1 9 17 25 33 41 48 kj i Frontal Vertical Horizontal
- 44. Tensor decomposition methods ● Approaches: – Tucker / HOSVD – CANDECOMP-PARAFAC (CP) – Hierarchical Tucker (HT) – Tensor-Train decomposition (TT) – NTF (Non-negative Tensor Factorization) – NTD (Non-negative Tucker Decomposition) – NCP (Non-negative CP Decomposition)
- 46. Tucker / HoSVD
- 47. CP ● The CP model is a special case of the Tucker model, where the core tensor is superdiagonal and the number of components in the factor matrices is the same. Solving by ALS (alternating least squares) framework
- 48. Background Model Estimation via Tensor Factorization https://github.com/andrewssobral/mtt/blob/master/tensor_demo_subtensors_ntf_hals.m
- 49. Background Subtraction via Tensor Decomposition https://github.com/andrewssobral/lrslibrary
- 50. Incremental Tensor Learning Interested in stream processing?
- 51. Incremental Tensor Subspace Learing
- 53. Incremental and Multifeature A total of 8 features are extracted: 1) red channel, 2) green channel, 3) blue channel, 4) gray-scale, 5) local binary patterns (LBP), 6) spatial gradients in horizontal direction, 7) spatial gradients in vertical direction, and 8) spatial gradients magnitude. values pixels features tensor model …
- 54. Incremental and Multifeature https://github.com/andrewssobral/imtsl Feature Extraction + iHoSVD Initialize Background Model frame model Foreground Detection Background Model Maintenance Low Rank Tensor Model Weighted combination of similarity measures x1 x3 x3 w1 w2 w3 yΣ ⁄
- 58. LRSLibrary The LRSLibrary provides a collection of low-rank and sparse decomposition algorithms in MATLAB. The library was designed for motion segmentation in videos, but it can be also used or adapted for other computer vision problems. Currently the LRSLibrary contains a total of 103 matrix-based and tensor-based algorithms. https://github.com/andrewssobral/lrslibrary
- 59. BGSLibrary The BGSLibrary provides an easy-to-use C++ framework based on OpenCV to perform background subtraction (BGS) in videos. The BGSLibrary compiles under Linux, Mac OS X and Windows. Currently the library offers 37 BGS algorithms.