An approach to detecting abnormal vehicle events in complex factors over highway surveillance video

The detection of abnormal vehicle events is a research hotspot in the analysis of highway surveillance video. Because of the complex factors, which include different conditions of weather, illumination, noise and so on, vehicle's feature extraction and abnormity detection become difficult. This paper proposes a Fast Constrained Delaunay Triangulation (FCDT) algorithm to replace complicated segmentation algorithms for multi-feature extraction. Based on the video frames segmented by FCDT, an improved algorithm is presented to estimate background self-adaptively. After the estimation, a multi-feature eigenvector is generated by Principal Component Analysis (PCA) in accordance with the static and motional features extracted through locating and tracking each vehicle. For abnormity detection, adaptive detection modeling of vehicle events (ADMVE) is presented, for which a semi-supervised Mixture of Gaussian Hidden Markov Model (MGHMM) is trained with the multi-feature eigenvectors from each video segment. The normal model is developed by supervised mode with manual labeling, and becomes more accurate via iterated adaptation. The abnormal models are trained through the adapted Bayesian learning with unsupervised mode. The paper also presents experiments using real video sequence to verify the proposed method.