Research On Online Object Tracking Algorithm Based On Sparse Representation

In recent years,the object tracking technology has attracted attention from researchers of the world and it has been applied to a variety of research prospects such as regional monitoring,traffic system,intelligent man-machine interface and so on.But there are many interference factors,such as occlusions,illumination change,pose change of the target,background interference in the sense of video sequences.At present,our research is focused on how to effectively improve the real-time performance and accuracy of tracking algorithms in practical application scenarios.Sparse representation based object tracking algorithm can better deal with changes in the appearance of the target.However,the dictionary used in this algorithm is easy to fail and result in tracking failure.In the thesis,we analyze sparse representation based tracking methods which was attractive in recent years and then proposes an improved tracking method on the basis of previous studies.The main contents of this thesis are as follows:Traditional sparse representation tracking use simple grayscale characteristics to calculate sparse coefficient,which is easily affected by the heavy occlusions and deformation.To this end,a object tracking algorithm fused by weighted local cosine and sparse representation is adopted.Firstly,the target template and the candidate targets are matched by the weighted local cosine similarity.The proposed local weighted algorithm increases the weights of the candidate targets which are not affected by occlusion,deformation,etc.Secondly,the target observation model makes use of the local information of the target by sparse coding and the dictionary is not updated.The construction of the reconstruction error considers the spatial layout between the local image blocks.The joint model considers the current state and the original state of the target,which improves the reliability of the observation model.Finally,the maximum posteriori probability is used to estimate the target state.Besides,a new update scheme is proposed to deal with target appearance changes problem.The experimental results show that the algorithm has strong robustness.L2-regularized least square method is adopted to deal with the high complexity in sparse representation models,Firstly,the extent of occlusion can be evaluated by L2 tracker.Secondly,convolutional networks is used to locate the target object if the extent of occlusion satisfies two inequality constraints.In order to make convolutional networks suitable for tracking tasks with high real-time requirements,this thesis uses a simple two-layer convolutional networks to represent the targets robustly.Finally,most of the insignificant samples are removed before applying convolutional networks,which reduces the complexity of the algorithm.The experimental results on numerous challenging image sequences show that the proposed method is more robust and stable than L2 tracker when the target object undergoes dramatic appearance changes such as pose variation or rotation and is superior in accuracy to other classical tracking algorithms.