Moving Target Tracking Based On Compressive Sensing

Target tracking is one of the key issue in the field of computer vision and the basis of many advanced computer vision applications,such as intelligent security,human-computer interaction,intelligent transportation,etc.As a highly efficient signal sampling method,the Compressive Sensing technology solves many of the computational problems caused by the high signal dimension,which significantly reduces the signal processing time and computational cost.The Fast Compressive Tracking algorithm successfully applied the Compressive Sensing theory to the reduced dimensionality of the high dimensional image feature,and achieved good tracking effect.This algorithm adopts a na?ve Bayesian classifier using immutable learning scheme so that algorithm cannot do some adaptive adjustments during to the variations of the tracking environment and the motion state of target object.To this end,this thesis has done research on learning the basic knowledge of the Compressive Sensing theory,understanding the commonly used tracking methods,focusing on making in-depth study of the Fast Compressive Tracking algorithm framework and theory,and made some improvements based on using the Compressive Sensing theory in target tracking.The research contents are listed as follow:(1)In order to enable the tracking algorithm to adaptively adjust the classifier to learn the feature information according to the change of the tracking environment and the movement state,this thesis improves the learning mechanism of the classifier update of the Fast Compressive Tracking algorithm,and proposes a new classifier adaptive learning update model,which makes the classifier more effective in learning the feature information and improves the classification and judgment ability of the classifier.In addition,a new strategy for judging the sample as the target location is proposed,which reduces the likelihood that the sample will be misclassified as target location by the classifier.(2)By combining the spatio-temporal context information into the Fast Compressive Tracking algorithm,the sample is weighted by calculating the distance between the predicted target position and the sampling sample in the continuous frame image,so as to effectively help the algorithm to better distinguish the target from the tracking background and improve the accuracy of the algorithm.Experiments on several datasets from Visual Tracking Benchmark have demonstrated that our algortithms have done effective improvements and received better tracking robustness.Although the time efficiency has declined,but still meet the requirements of real-time tracking.