Robust Visual Tracking via Incremental Subspace Learning

and Local Sparse Representation

ABSTRACT:

Single target tracking is an important part of computer vision, and its robustness is always restricted by target occlusion, illumination change, target pose change and so far. To deal with this problem, this paper proposed a robust visual tracking based on incremental subspace learning and local sparse representation. The algorithm adopts local sparse representation to test occlusion and rectifies the incremental learning error according to the occlusion detection outcome and to overcome the influence of occlusion on target template. Moreover, similarity between target templates and candidate templates is computed on the basis of local sparse representation. In the frame of particle filter, target tracking is achieved by combining incremental error and similarity measurement. The experimental resulting in several challenging sequences shows that the proposed method has better performance than that of state-of-the-art tracker.

PROJECT OUTPUT VIDEO:

EXISTING SYSTEM:

Current algorithms of visual tracking can be classified into two categories: generative methods and discriminative methods. The generative methods are mainly to find the most similar area with target template, which are based on appearance model. This method is built on target template and subspace model. Besides, methods based on super pixels model, color model, coupling layer model and part idea model are also used to model the target appearance which captures new appearance change by template update.

The discriminative methods transform tracking problem into binary classification problem, i.e., extracting target from background. These methods extract features from target and background to build training set, and accuracy of tracking often depends on the classifier.

Zhong et al. proposed a sparse corporation appearance model, in which tracking results are identified by corporation of a sparse discriminative classifier (SDC) and a sparse generative model (SGM). SDC extracts tracking target by using global templates; as to SGM, spatial information of local block is used to measure the similarity of candidate and target.

DISADVANTAGES OF EXISTING SYSTEM:

Unsupervised learning is used to train classifiers, but the tracking result has low accuracy for lack of supervision mechanism for the classifier update. Semi-supervised and multiple-instance learning are proposed to deal with this problem.

PROPOSED SYSTEM:

The main contributions of this paper are listed as follows:

1. The local sparse representation is used to detect occlusion in the tracking target which can effectively reduce the impact of occlusion on tracking performance. The method of combining global information and local information is adopted in the decision-making link to increase the applicability of the algorithm.

2. When the template is updated, the occlusion information is being fully used to prevent occlusion from being updated to the target template.

ADVANTAGES OF PROPOSED SYSTEM:

The algorithm used incremental subspace model to get the feature vector and the mean of target template and then calculated the incremental subspace of target template.

The incremental error of the candidate target was obtained by the projection of the candidate target on the incremental subspace. In order to overcome the influence of target occlusion, illumination change and scale change on target tracking, we proposed occlusion detection by using local sparse representation and corrected incremental learning error according to the candidate target occlusion.

MODULES:

1. Video Sequence Acquisition

2. Visual Tracking

MODULE DESCRIPTION:

1. Video Sequence Acquisition:

• Image acquisition in image processing can be broadly defined as the action of retrieving an image from some source, usually a hardware-based source, so it can be passed through whatever processes need to occur afterward.

• Performing image acquisition in image processing is always the first step in the workflow sequence because, without an image, no processing is possible. The image that is acquired is completely unprocessed and is the result of whatever hardware was used to generate it, which can be very important in some fields to have a consistent baseline from which to work.

• In this stage, test video sequence acquire from public database.

2. Visual Tracking:

• In this paper, the algorithm is based on particle filter as a whole frame, using incremental subspace learning and local sparse representation to calculate similarity of candidate targets and target template.

• The similarity is used as observation likelihood of candidate to predict the optimal statement of tracking target of next frame and in specified way, and target tracking is achieved.

Main contributions of the work:

• The local sparse representation is used to detect occlusion in the tracking target which can effectively reduce the impact of occlusion on tracking performance.

• The method of combining global information and local information is adopted in the decision-making link to increase the applicability of the algorithm.

• When the template is updated, the occlusion information is being fully used to prevent occlusion from being updated to the target template.

SYSTEM REQUIREMENTS:

HARDWARE REQUIREMENTS:

• System          : Pentium Dual Core.

• Hard Disk      : 120 GB.

• Monitor          : 15’’ LED

• Input Devices : Keyboard, Mouse

• Ram                : 1GB.

SOFTWARE REQUIREMENTS:

• Operating system : Windows 7.

• Coding Language : MATLAB

• Tool                       : MATLAB R2013A

REFERENCE:

Guoliang Yang, Zhengwei Hu, Jun Tang, “Robust Visual Tracking via Incremental Subspace Learning and Local Sparse Representation”, SPRINGER 2017.