Object Tracking Based On Sparse Representation And Its Implementation With CUDA

Visual object tracking techniques has been widely applied in many real time scenario such as Robot Navigation,Intelligent Surveillance,etc.,due to the improving computation capability of modern hardware and the theoretical development of computer vision in recent decade.With new applications arrive,new challenges emerge.Facing the more and more complex object appearance,object tracking technology has to develop sophisticated coping strategies and acquire ideas from other research areas.Sparse representation was initially applied to image processing tasks such as image denoising,image up-scaling.Then it is applied to visual object tracking techniques,and gained extraordinary performance due to its ability to adapt to complex appearance changes.In visual object tracking task,Sparse representation is usually combined with particle filter framework.The sparsity and layout of decomposition coefficients can be used to construct a robust likelihood function,in order to decrease error due to local occlusions,illumination changes.The time complexity for computing sparse codes is usually high because iterative methods are involved.This is one of substantial obstacle to real time application such as visual object tracking.GPU has been widely used to solve computationally intensive tasks,and result in massive performance boosting.CUDA toolkit provides an efficient way to apply GPU to general-purpose computation.This thesis implements sparse representation based object tracking algorithm on NVIDIA GPU using CUDA.SpaRSA is used compute the sparse codes of each image patch,and is implemented using CUDA kernel function and CUBLAS library.Alignment-pooling method is used to compute likelihood of each candidate region.The subspace of target images is updated via Sequential Karhunen–Loeve(SKL)algorithm.The dictionary of local patches is then constructed and updated using the latest result and the subspace of target images.Matrix computation in SKL is implemented via CUBLAS and CUSOLVER libraries.CUDA kernel functions are used to other matrix and vector operations.The primary portion of computation are completely implemented on GPU.The participation of CPU are limited as possible.The processing speeds on test sequences are approximately 50 ~ 60 fps(512 particles)and 80 ~ 100 fps(256 particles).Experiments shows that this implementation has gain robust tracking result and good acceleration rate.The SpaRSA kernel function takes 95.8% of the processing time,and reaches 97.2% device occupancy while theoretical occupancy is 98.4%.