Target Tracking Method Based On The Covariance Matrix

Object tracking is one of the core research topics in computer vision. It haswidespread application prospect in motion-based recognition, intelligent traffic,automatic navigation, military monitor tracking, video security monitoring, biomedicalimage analysis and so on. This paper mainly studies various object tracking algorithmsbased on covariance matrix.A compact covariance matrix is proposed to model objects in classic covariancetracking method which fuses various kinds of features conveniently. However, it can'tachieve real-time tracking if the object is big and its presentation capability is limited.Aimed at the former problem, CUDA based parallel tracking method is proposed. First,image features are computed and integral images are constructed in parallel. Then,covariance matrices are calculated in parallel. Finally, the distance metric is developedin parallel. Compared with the efficient method on CPU, the speedup ratio is over10.As to the latter problem, this paper analyzes and makes improvements fromdifferent viewpoints to enhance tracking performance. Classic method ignores imagenoises and background. This paper proposes orientation covariance tracking method toaddress the problem. A spatial kernel is proposed to weaken background and asmoothing kernel to reduce errors of orientation angles. Then, the angle space ispartitioned into several subspaces. Finally, a steepest descent algorithm is performed fortracking. Classic method ignores the distribution of data. This paper proposes Gaussianmixture model based tracking method to address the problem. A neighboring covariancematrix is extracted of each pixel. Then, it is mapped to Euclidean space based on Liegroup and Lie algebra as training data to estimate a Gaussian mixture model. Finally, athreshold is used to detect the object. Classic method ignores spatial information andintrinsic correlation of data. Tensor based covariance tracking method is proposed toaddress the problem. The data are unfolded to three modes, K-L transform is used to getthe principle components. Then, the reduced third-order tensor is used to model objects.Finally, an efficient incremental update mechanism is proposed. Experiments show thatthe three methods all achieve better performance and higher tracking efficiency.