Moving Object Detection in Complex Scene Using Spatiotemporal Structured-Sparse RPCA
Moving object detection is a fundamental step invarious computer vision applications. Robust Principal ComponentAnalysis (RPCA) based methods have often been employedfor this task. However, the performance of these methodsdeteriorates in the presence of dynamic background scenes,camera jitter, camouflaged moving objects, and/or variations inillumination. It is because of an underlying assumption that theelements in the sparse component are mutually independent, andthus the spatiotemporal structure of the moving objects is lost. Toaddress this issue, we propose a spatiotemporal structured sparseRPCA algorithm for moving objects detection, where we impose spatial and temporal regularization on the sparse component inthe form of graph Laplacians. Each Palladian corresponds toa multi-feature graph constructed over super pixels in the inputmatrix. We enforce the sparse component to act as eigenvectorsof the spatial and temporal graph Laplacians while minimizing the RPCA objective function. These constraints incorporate aspatiotemporal subspace structure within the sparse component.Thus, we obtain a novel objective function for separating movingobjects in the presence of complex backgrounds. The proposedobjective function is solved using a linear zed alternating directionmethod of multipliers based batch optimization. Moreover,we also propose an online optimization algorithm for real-timeapplications. We evaluated both the batch and online solutionsusing six publicly available datasets that included most of theaforementioned challenges. Our experiments demonstrated the superior performance of the proposed algorithms compared withthe current state-of-the-art methods.
MOVING Object Detection (MOD) is an important step in many computer vision applications, such as video surveillance , anomaly detection , salient object detection , video rain drops removal , video inpainting , visual object tracking , and human action recognition . MOD mainly aims to separate the foreground objects from a background scene, however MOD becomes challenging when a scene contains dynamic backgrounds comprising variations in illumination, swaying bushes, rippling water surfaces, the performance of MOD methods is also degraded by sequences containing moving objects in all of the frames as a bootstrapping effect, shadows caused by moving objects, and jitter-induced motion in the background. Many approaches have been reported for handling complex background scenes
The eigenvector correspondingto the minimum non-zero Eigenvalue is also known as Fiedlervector and it defines two partitions of the graph based on thesigns of its coefficients. We enforce the resulting sparse matrixto act as the eigenvectors matrix of the spatial and temporal graph Laplacians. By incorporating these spectral clusteringbased objective functions into the low-rank decomposition, weensure that the resulting sparse matrix encodes the spatial and
temporal connectivity at the super pixel level. Incorporatingspatiotemporal graph Palladian matrices into objective functionallows the proposed SSSR algorithm to detect movingobjects in a more robust manner in complex scenes, evenwhen the appearance of the moving objects is similar to thebackground (Fig.1, column d), which is difficult to handleusing existing methods.
Some RPCA enhancements have been proposed to improvethe sparsity patterns of the moving objects. Zhou et al. proposed DECOLOR, which incorporatesMarkov random field constraints into the sparse matrix F. Xinet al. proposed the GFL method, which encodes the fusedlasso regularization in the sparse component. The smoothness effect of the Markov random field and fused lasso effectivelyeliminates the noise and small background movements (dynamicbackground pixels). However, the foreground regionstend to be over-smoothed to an undesirable extent becauseof the strict smoothing constraints. the over-smoothing degrades the MOD performancedue to the incomplete foreground or the merging of distinctmoving objects into one segment. The background regionwithin distinct moving objects is detected as part of the moving object segment.
Moving Object Detection Using Spatiotemporal Structured-Sparse RPCA
The smoothnesseffect of the Markov random field and fused lasso effectivelyeliminates the noise and small background movements (dynamicbackground pixels). However, the foreground regionstend to be over-smoothed to an undesirable extent becauseof the strict smoothing constraints.
We use a super pixel representation in the SSSR algorithm. Afeature matrix D is computed using super pixels and multiplerepresentations. The spatiotemporal structure of the sparsecomponent is preserved by encoding the pair wise similaritiesamong the columns and rows of the matrix D. For this purpose,we construct two graphs one among the temporal super pixelsor columns of matrix D, and the second among the spatiallocations or rows of matrix D.
Most of the traditional methods process the input datamatrix by batch/offline processing, which requires that a largenumber of frames are stored in memory. Batch processingobtains good accuracy but it may cause delays and theinput frames might not be available for real-time scenarios;therefore, some online methods have also been proposed inliterature.
Used for batch processing while anonline solution is designed to handle the real-time processingchallenges. Our proposed online solution significantly speedsup the process and the MOD method is suitable for real-timeapplications. Initial results of this study are recently reported .The current study is a significant extension of both theoretically and experimentally.
The conventional RPCA model given by lacks structuralinformation therefore it is not practical for complex backgroundscenes. We propose to handle this problem more effectivelyby enforcing spatial and temporal structural constraintson the sparse matrix F.
Model is essentially a convex optimization problem, whichcannot be solved directly using the conventional ADMMmethod since contains more than two blocks of variables( e.g., B, F, S, and H). To solve the proposed model givenby , we formulate batch optimization using a liberalizedversion of ADMM method because of its efficiency .
We overcome this problem by solving model using aniterative or online optimization scheme, where we first convertinto unconstrained problem.
Toaddress this problem, we use a basis update scheme based onstochastic gradient descent optimization. Given the projectedcolumn, the gradient with respect to basis U can be derived.
- MATLAB 7.5 and above versions
0.00 average based on 0 ratings
More Things You Might Like This
Abstract: Although the educational level of the Portuguese population has improved in the last decades, the statistics keep Portugal at Europe’s tail end due to its high student failure rates. In particular, lack of success in the core classes of Mathematics and the Portuguese language is extremely serious. On the other hand, the fields of
Abstract: Advances in natural language processing (NLP) and educational technology, as well as the availability of unprecedented amounts of educationally-relevant text and speech data, have led to an increasing interest in using NLP to address the needs of teachers and students. Educational applications differ in many ways, however, from the types of applications for which
Machine Learning based Regression Model for Prediction of Soil Surface Humidity over Moderately Vegetated Fields
Abstract: Agriculture is one of the major revenue producing sectors of India and a source of survival. Numerous seasonal, economic and biological patterns influence the crop production but unpredictable changes in these patterns lead to a great loss to farmers. These risks can be reduced when suitable approaches are employed on data related to soil