Pedestrian Classification Based on Full-SVM Decision Tree

Abstract

Visual analysis has potential to be used for recognition, and it is one of the hottest but most difficult subjects in computer vision. In order to identify pedestrian movement in an Intelligent Security Monitoring System, the video activity in the prospect is represented by a series of spatio-temporal interest points. Since human posture has the characteristics of uncertainty and illegibility, the clustering centers of each class are computed by fuzzy clustering techniques. We presented a pedestrian classification method based on improved support vector machines in order to solve non-rigid objects that are difficult to identify in intelligent monitoring systems. The Support Vector Machine technology and the decision tree have combined into one multi-class classifier so as to solve multi-class classification problems. Then a full-SVM (Support Vector Machine) decision tree is constructed based on the conventional decision tree. At last, the method is evaluated on the KTH action dataset and receives a comparatively high correct recognition rate.