Research On Video Object Tracking Based On Compressive Sensing

As the research hotspot and difficulty in the computer vision field,the object tracking has been widely applied in the intelligent human-computer interaction,visual navigation,intelligent video surveillance and other fields.However,it is still an extremely challenging subject to develop a kind of object tracking algorithm which is able to achieve the stable,fast and high-precision tracking at various kinds of complex scenes.In recent years,the compressive tracking algorithm has become the research hotspot because of its good tracking performance.In order to improve the tracking performance of compressive tracking algorithm under the complex scene,this paper conducts research.The main research achievements of this paper are as follows:In order to improve the tracking stability and accuracy of the compressive tracking algorithm under illumination and occlusion,a compressive tracking algorithm combining online feature selection with covariance matrix is proposed.Firstly,introduce the online feature selection based on the Hellinger distance at the feature extraction phase,and dynamically choose the feature with high confidence score in the feature pool to build the Bayes classifier.And then,integrate the covariance matrix under the compressive tracking frame to enhance the representation ability of the algorithm to the object,combine the Haar-like feature with covariance matrix to build the object model,and take the candidate sample corresponding to the maximum response value as the tracking result.Finally,optimize the updating way of classifier parameter,to adaptively update the classifier parameter according to the similarity of object template and tracking result.Aiming at the problem of compressive tracking algorithm can not adapt to the change of the object scale and the sample weight is not considered,a kind of compressive tracking algorithm based on the particle filter and sample weighting is presented.Firstly,improve the compressive feature in the original compressive tracking algorithm,to extract the compressive feature with scale invariance for building the object appearance model.And then,introduce the thought of sample weighting,give different weights to the positive sample in accordance with the distance difference between the positive sample and object,and improve the classification precision of the classifier.Finally,make the dynamic state estimation byintegrating the normalized rectangle feature under the particle filter frame,make estimate and prediction for the particle state by utilizing a second-order autoregressive model at the particle prediction phase,update the particle state by virtue of observation model,and resample the particles to avoid the particle degeneracy.The experimental results show that compressive tracking algorithm combining feature online selection with covariance matrix is of higher tracking robustness and accuracy under illumination and occlusion.Compressive tracking algorithm based on the particle filter and sample weighting can adapt to the change of object scale,with higher tracking stability and accuracy.