Pedestrian Behavior Anomaly Detection Based on Dynamic Mode Decomposition and One-Class SVM

Anomaly detection of human behavior, which has been widely applied in video supervision in recent years, aims at recognizing the human behaviors that are out of the normal scope in high-level semantic. With the development of deep learning (DL), many DL-based methods for behavior recognition have been developed. However, the lack of negative samples hinders the application of DL-based methods. Thus, this work proposed a new framework based on dynamic mode decomposition (DMD)which is free from the requirements of negative samples. First, key points of the human body in each frame of a video are exacted by the -deep neural networks (DNN)which have been solved in recent studies. Second, the key points are projected into a mode space of much lower dimensions with DMD. At last, the detection results are obtained with One-class SVM, which is a classification method that does not require negative samples. The data set we use in the experiment is the CMU motion data set. Experiment show that the accuracy of DMD combined with support vector classification (SVC) achieve 85% while the accuracy of DMD combined with one-class SVM achieve 81%, and the proposed framework can distinguish anomaly behavior with a small number of samples.