Research On Long-term Robust Infrared Object Tracking

Infrared object tracking has drawn much attention in the field of computer vision.At present,it has been widely used in military and civil fields such as precision guidance,reconnaissance and video surveillance.However,due to the low contrast and lack of texture and color information of infrared images,the existing visual tracking algorithms are difficult to apply into infrared object tracking.In this paper,a generative tracking algorithm is proposed,which generates feature vector by sparse coding,and locates object under the framework of particle filtering.The proposed algorithm can achieve long-term robust tracking.The main contributions of this paper are as follows:Firstly,a feature extraction method based on sparse coding is proposed to characterize the object.The object region is divided into several overlapping image blocks,and the SIFT feature is extracted from each block to solve the problem of occlusion,rotation and other issues.the SIFT feature of each image block is sparsely encoded online to describe the image more accurately.In which,the over-complete dictionary is trained offline by K-SVD method.At last,based on the spatial pyramid matching technology,the sparse vectors of all the image blocks are multiscale spatial max pooled to get a fixed-length output vector as the object feature vector.Thence,the global and local information of the image is preserved.Therefore,the tracking algorithm can effectively solve occlusion problem and improve tracking robustness.Secondly,in order to improve the real-time performance of the algorithm,an optimization modeling method is proposed.A SIFT-like fast feature extraction method is adopted to obtain the local feature descriptors of each image block,which avoids the cumbersome operation of down-sampling and interpolation process.And locality-constrained linear coding is used instead of sparse coding,which uses the locality constraints to project each local feature descriptor into its local-coordinate system.Compared with sparse coding,locality-constrained linear coding can be performed very quickly for appearance modelling because it has an analytical solution derived by a three-step matrix calculation.Hence,it can reduce the computational complexity and improving the efficiency of the algorithm.Thirdly,a scale adaptive method is implemented under the framework of particle filtering.Different from most tracking algorithms using scale pyramid or scaling factor,scale variation in both horizontal and vertical directions of the particle is given in order to avoid repeated calculation of same particle.The Markov prediction model accurately predicts the range of scale changes,and updates the object's size adaptively.In addition,the motion information is taken into account,so as to solve the problem that the traditional tracking algorithm cannot track the fast-moving object or similar moving objects interfering with each other.Finally,in order to demonstrate the tracking effect of the algorithm and evaluate the influence of each input parameter,a Matlab GUI platform is also developed in this paper.The GUI can read and save files,input parameters and real-time display and show the tracking results.Users can select the video sequence through the controls on the interface and set the parameters.Quantitative experiments conducted on this experimental platform,demonstrate that our proposed algorithm performs favorably against the other eight state-of-the-art trackers on 10 challenging test sequences.