Research On Scale Adaptive Object Tracking Algorithm With Compressed Sensing

The video object tracking technique is one of important research direction in the field of computer vision,which has been widely applied to in the aspect of intelligent human-computer interaction detection,military guidance.The task of object tracking is to find the object location in video sequences,and then analysis the object behavior deeply.The object tracking technique has achieved rapidly development in recent years,and a lot of research achievements have been applied to practical field.Due to the influences of illumination change,occlusion,background interference and rotation,object tracking is still a challenging subject.Recently,compressive sensing(CS)theory has been a popular advanced theory that attracts more and more researchers pay attention to.Sparse representation has become a new hotpot research and has obtained many research results in object recognition and object tracking.Some improved tracking algorithms based on the traditional compressive tracking(CT)algorithm were proposed.In this paper,the main research contents and innovation points are as follows:(1)In this paper,the feature extraction measures and compressive sensing theory were studied,which were applied to object tracking algorithm.Meanwhile some mainstream object tracking methods which were chosen from generative model and discriminative model were chosen to improve.Finally,the improved tracking algorithms were compared with the other tracking algorithms by experiments.(2)A feature weighting compressed object tracking algorithm is proposed.The algorithm extracts feature from four sub-regions partitioned from the global region by measurement matrix.The distribution of regional features was estimated and the weights were calculated based on the classification results to obtain the object's position in next frame.According to the similarity of the original frame and current frame to update parameters of classifier adaptively.The algorithm improves the performance of anti-occlusion and holds robustness in the process of object tracking.(3)A multi-feature fusion compressed object tracking algorithm is proposed.Firstly,the color features and texture features were used to describe the object.Then random measurement matrix was used to compress the texture feature and color feature of object,the object model was set up and the object state was estimated based on the principle of particle filter to obtain object's position.Aim to adapt to the change of object and environment,an adaptive updating strategy of object template was designed,and number of particles was adjusted dynamically.The experimental results show that the algorithm can track the object accurately under the complex environment of partial occlusion and illumination changes.(4)A scale adaptive compressed object tracking algorithm is proposed.The proposed algorithm extracts the low-dimensional gray and texture features of the object and its surrounding region by sparse measurement matrix which reduces the computation complexity.Then the tracking task is formulated as a binary classification via support vector machine(SVM)classifier with online update in the compressed domain.Meanwhile utilize the classifier to obtain the object's position in new frame.In addition,build the Hamming distance between hash values of current object and original one to match the template to achieve adaptive template size.Numerous experimental results show the proposed algorithm can keep tracking effectively when the tracked object is under the situation of scale variation.