Abnormal Crowd Motion Detection with Hidden Markov Model

Abstract

stations,etc. With the increasing demand of surveillance of various human activities, an efficient automated surveillance system to detect anomalies has become important. There is a survey on visual surveillance in [1], and a lot of problems have not resolved in surveillance applications nowadays as discussed in some papers [2]. Crowd feature extraction and crowd modeling are two important approaches of crowd analysis in Video surveillance. Most detection methods use motion features such as image gradient, texture information, optical flow, or spatial-temporal volume characteristics [3-4]. These features allow us to analyze dense crowds using characteristics of a crowd rather than individuals. Some works in the analysis of crowds usually assume that individuals can be tracked and identified in the crowd [5]. For example, the method proposed in [6] combines the inter-images difference based on entropy image and optical flow computed by a local method with a hierarchical coarse-to-fine optical flow estimation, and this method has a big computation load. Once motion features are extracted, another approach is to model them to represent normal motion behaviors. Abnormal crowd motion is defined as a sudden change or irregularity in a crowd motion. Suppose there exists a mathematical model describing the normal behaviors in the crowd videos, anomaly detection is done by modeling normal behaviors, and a crowd behavior is declared as anomaly if its characteristic does not comply with the learning model. Many approaches use parametric models, however, in parametric approach the data characteristics have to be approximated with a standard distribution and in many cases such approximations do not work well. In [7], Social Force model is used to detect and localize abnormal behaviors in crowd videos by Mehran et al. Their result shows the approach has a better performance than similar approach based on pure optical flow. Andrade et al. [8] combined spectral clustering, Principal Components Analysis (PCA) and Hidden Markov Model (HMM) to detect the crowd emergency scenarios, but the eigenvectors obtained by dimensionality reduction with PCA on the optical flow fields can not adequately reflect the motion, and this approach was only tested in simulated data. Our approach derives variations of motion patterns though direction distribution of the crowd motion obtained by optical flow and these variations are encoded with HMM Abnormal Crowd Motion Detection with Hidden Markov Model Dongping Zhang, Yafei Lu, Xinghao Jiang, Huailiang Peng