Robust Infrared Visual Object Tracking Based On Improved Staple Algorithm

Current infrared visual object tracking algorithms are not able to track objects efficiently when faced with multiple complicated situations.Staple is an algorithm which makes full use of information of objects aiming at visible image primarily.There exits many differences between infrared image and visible image,giving rise to unsatisfactory performance when Staple is applied to infrared image.This paper analyzes the reason why this phenomenon is produced and proposes corresponding improvement based on Staple algorithm.The details are as follows:(1)In order to solve the problem that infrared imagery is short of information,employing histogram of sparse codes as feature to represent object is proposed,which takes advantage of structural information of infrared objects to enhance representative ability of this tracker.Staple algorithm calculates template score based on histogram of oriented gradient which relies on gradient calculation,thus it's difficult to extract features with sufficiently discriminative representation.Histogram of sparse codes is independent of gradient,meanwhile it captures edges and points feature of objects.By setting learning rates of two correlation filters using two different features to different values the purpose of maintaining accuracy and stability of proposed trackers is achieved.(2)In the light of stability of brightness information of infrared image,gray histogram score is redesigned by utilizing distractor aware model to decrease negative influence from distractors in surroundings.Distractor aware model score is obtained by modeling targetbackground and target-distractor at the same time and merging both models reasonably.Then gray histogram score based on distractor aware model and template scores based on HOG and HSC are merged in a certain way,making full use of gray information of infrared image which is vital for infrared visual tracking task.(3)Adaptive spatial windowing is employed to substitute cosine window to weight features.Spatial windowing endows obvious features with higher weights and alleviates boundary effect,rendering superior capability to deal with deformation and partial occlusion to correlation filter.Moreover,inspired by intrinsic relationship between structural information and gray information,this paper proposes to integrate spatial windowing and distractor aware model to endow pixels belonging to the target with higher weights.At the same time a spatial windowing threshold is used to guarantee that the proposed algorithm is able to position target precisely when abrupt motion or deformation occurs.(4)Parameters in Staple algorithm update every frame regardless of quality of training samples in each frame.A novel judgment standard——information entropy energy is pro-posed as feedback,which is used to estimate whether the target in current frame can be treated as new training sample by measuring concussion severity of response map of correlation filter.The greater the value of information entropy energy is,the more severe the concussion is,and when information entropy energy is greater than preset threshold the parameters of tracker will not update and vice versa.