Detecting anomaly based on time dependence for large scenes

We propose a novel approach for the crowd anomaly detection in multiple cameras with non-overlapping and visible views. As we all know that there are some kinds of information hidden in the non-overlapping fields always. In this paper, we will mine time dependence data so that we can analyze the crowd anomaly detection from time dimension's angle. Firstly, we have to preprocess the real scene using optical flow. Secondly, we build a model of crowd movement. We build the model of crowd movement using random data of the simulated scene and real scene and based on neighborhood weighted fuzzy c-means(NW-FCM) algorithm. Thirdly, we analyze local and global path based on time dependence data. We research probability of one trajectory through one piece of local path. Then we study the global path based on the Bayesian Information Criterion (BIC) and Markov Chain Monte Carlo (MCMC). At last, we can analyze the crowd anomaly detection. There are two kinds of anomaly events including abnormal retention events and abnormal moving event. We set up the empirical threshold value of probability P
_e
. If the probability of detected model is less than P
_e
, the detected model is marked as the crowd anomaly. We judge the detection system based on the confusion matrix. The global comprehensive assessment criteria for the real scene is 95.6%. Experimental results show the anomaly detection is precise.