Research On Object Tracking And Point Set Registration Based On Fuzzy Support Vector Machine

Computer vision,as a vital component of artificial intelligence,has developed rapidly in recent years.Due to its excellent performance,support vector machine has been widely used in various tasks of computer vision.Samples are contaminated or the requirement of some tasks making different samples have different contributions to the learning result.The limitation of the design of support vector machine makes it only possible to treat all samples equally.So,fuzzy support vector machine that introduces a fuzzy membership for each sample to describe the importance of the sample to the learning result has begun to receive extensive attention.In this paper,we investigate the fuzzy support vector machine,and propose a new fuzzy support vector regression,and then research on object tracking and point set registration by means of the fuzzy support vector machine and proposed fuzzy support vector regression,respectively.The main research contents and innovations are as follows:1.We use the coordinate descent method to solve the fuzzy support vector machine with a squared hinge loss.We propose a vector-valued fuzzy support vector regression with a quadratic ε-insensitive loss and manifold regularization,and use the coordinate descent method to solve it.2.To pay different degrees of attention to object and its surroundings,we propose a tracker that is built by means of the fuzzy support vector machine with a squared hinge loss.We use the coordinate descent method to update the tracker parameters online,and employ dense sampling and correlation filter to realize the high efficiency of sampling and calculation,so that the tracker can achieve high-speed tracking.The tracker is then generalized to obtain its multi-channel variant,kernelized variant and scale-adaptive kernelized variant.Extensive experiments show that the proposed scaleadaptive and kernelized tracker has higher tracking accuracy than other trackers,and the speed is close to real-time.3.To learning from true matches,we propose a transformation model that is built by means of the vector-valued fuzzy support vector regression with a quadraticε-insensitive loss and manifold regularization,and then we propose a new method for nonrigid point set registration,which is based on finding feature matches and estimating the transformation function.We introduce a mixture model to assess the probability that a pair of points is a true match.The quadratic ε-insensitive loss is insensitive to noise,and the regularization forces the transformation function to preserve the intrinsic geometry of the input data.We use the coordinate descent method to solve the transformation model and accelerate its speed by means of a sparse approximation.Extensive experiments show that our approach is efficient and outperforms other methods.