A simplified minimum enclosing ball based fast incremental support vector machine (SVM) algorithm for person detection and tracking

摘要

In order to meet the requirements of stable person detection and tracking techniques in dynamic visual system, we propose a simplified minimum enclosing ball based fast incremental support vector machine (SVM) algorithm for person detection and tracking. Based on the simplified minimum enclosing ball (MEB) method, we propose a simplified and fast incremental algorithm to compute the MEB. By utilizing the equivalence between MEB and the dual problem in SVM, we achieve the online and incremental adjustment of the SVM classifier coefficients. The proposed method do not need to solve the quadratic programming problem. It is fast for training. Moreover, it can achieve the online update of classifiers for object tracking with small sample size. Finally, the efficiency of the proposed incremental SVM is validated by detection experiments on dynamic pedestrians tracking system.