Research On Target Tracking Algorithm Based On Compressed Sensing Theory

As an important part of computer vision and image processing research,visual target tracking technology is widely used in many fields.However,with the complexity and diversification of application scenarios and the accuracy of tracking performance requirements,how to achieve high efficiency in challenging conditions such as target deformation,occlusion,fast motion and low background similarity,dramatic changes in ambient illumination,narrow viewing angles and even more complex conditions.Real-time,accurate and robust tracking is a difficult part of target tracking technology.The current research on the hot tracking algorithm based on deep feature representation is complicated by the complexity of the network hierarchy,and the computational processing overhead is huge.At the same time,a large number of training data sets are needed,which cannot be applied to fields with weak computing power such as embedded system.The tracking algorithm based on traditional shallow feature extraction and template matching still has important research significance.This paper mainly studies the tracking problem of moving single target,starting from the traditional feature compression tracking algorithm,aiming at poor tracking robustness due to target motion deformation,template redundancy and computational overhead due to blindness of positive and negative sample search sampling.And so on,the corresponding tracking algorithm is proposed.The main contents of the research include the following two aspects.1.A feature compression tracking algorithm based on enhanced scale estimation is proposed.In the target tracking process,more background noise is introduced into the scale variation of the target,but the sampling is reduced when the target scale becomes small,resulting in weak algorithm robustness.Problem:To achieve robust tracking of visual targets in complex background environnents,a discriminant correlation filter is separately set for scale estimation,online learning updates sample scales,real-time matching of optimal target sizes and updating of feature sample block sizes,compression of sample features,reduce dimensionality and learn online to update classifier parameters,reduce computational overhead,and improve tracking robustness.2.An directional sampling tracking algorithm based on Kalman filter model for target motion trajectory prediction strategy is proposed.By limiting the range of positive and negative sample sampling fields of foreground and background,the tracking algorithm is greatly reduced for candidate target samples and background noise samples.The number solves the high sampling redundancy problem of the traditional feature compression tracking algorithm.By appropriately increasing the sampling rate of positive and negative samples in the constrained sampling domain,the real-time performance and tracking accuracy of the algorithm are doubled.