Statistical Hypothesis Detector for Abnormal Event Detection in Crowded Scenes
Statistical Hypothesis Detector for Abnormal Event Detection in Crowded Scenes Abnormal behavior detectio. A robust and efficient anomaly detection technique is proposed, capable of dealing with crowded scenes. Many existing methods learn a normal event model in the training phase, and events which cannot be well represented are treated as abnormalities. However, they fail to make use of abnormal event patterns, which are elements to comprise abnormal events. Moreover, normal patterns in testing videos may be divergent from training ones, due to the existence of abnormalities.
Multifeatures based on trajectories of moving objects, such as mean value, speed, and acceleration. Each feature is treated as a distinct space, and is applied with a clustering algorithm. The final clustering results are obtained by considering clusters from all feature spaces. Clusters with few trajectories and trajectories far away from the cluster centers are treated as abnormalities. Cheng and Hwang make use of an adaptive particle sampling and the Kalman filtering to deal with occlusion and object segmentation problems, and obtain reliable trajectories. After that, abnormal trajectories are recognized through classification. Besides the object level trajectory extraction, tracking at particle and feature point level have also been taken into consideration.
Proposed a Lagrangian particle dynamics approach, and extract chaotic invariant features from representative trajectories. Normal motion patterns are modeled with a probabilistic framework, and abnormal events are detected with a maximum likelihood estimation criterion. Cui et al. Tracked interest points and calculate interaction energy potentials to represent the crowd dynamic. By analyzing features of different patterns, abnormalities are judged by a trained classifier. Although trajectory-based features are high-level semantic, they are no longer effective when the density of a crowd increases. This is because of the unreliable tracking under conditions of inevitable overlaps and occlusions.
Represented motion patterns with a multiscale HOF. By calculating sparse representation coefficients with a trained dictionary, abnormal events are detected as samples with large reconstruction costs. Adam et . Adopted histograms to model the probability of optical flow at a group of fixed spatial locations. Kim and Grauman used a mixture of probability principle component analyzers (MPPCAs) to model local optical flow, and account for space-time interactions with a Markov random field. Mehran et al. analyzed crowd behaviors based on a social force (SF) model, where the interaction forces are calculated with optical flow.Although trajectory-based features are high-level semantic, they are no longer effective when the density of a crowd increases. This is because of the unreliable tracking under conditions of inevitable overlaps and occlusions.