2018/02/09 by Michael Ying Yang, Yang, Michael Ying, Wentong Liao +5
Computer Science · #Anomaly Detection Techniques and Applications #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Human Pose and Action Recognition #Video Surveillance and Tracking Methods
paper · pdf · doi:10.48550/arxiv.1802.03257
openalex publication_date 2018/02/09 · openalex created_date 2022/09/28 · openalex updated_date 2026/07/28
In this paper, we present an unsupervised learning framework for analyzing\nactivities and interactions in surveillance videos. In our framework, three\nlevels of video events are connected by Hierarchical Dirichlet Process (HDP)\nmodel: low-level visual features, simple atomic activities, and multi-agent\ninteractions. Atomic activities are represented as distribution of low-level\nfeatures, while complicated interactions are represented as distribution of\natomic activities. This learning process is unsupervised. Given a training\nvideo sequence, low-level visual features are extracted based on optic flow and\nthen clustered into different atomic activities and video clips are clustered\ninto different interactions. The HDP model automatically decide the number of\nclusters, i.e. the categories of atomic activities and interactions. Based on\nthe learned atomic activities and interactions, a training dataset is generated\nto train the Gaussian Process (GP) classifier. Then the trained GP models work\nin newly captured video to classify interactions and detect abnormal events in\nreal time. Furthermore, the temporal dependencies between video events learned\nby HDP-Hidden Markov Models (HMM) are effectively integrated into GP classifier\nto enhance the accuracy of the classification in newly captured videos. Our\nframework couples the benefits of the generative model (HDP) with the\ndiscriminant model (GP). We provide detailed experiments showing that our\nframework enjoys favorable performance in video event classification in\nreal-time in a crowded traffic scene.\n