Abnormal Event Detection Algorithms And Implementation

In this article,we concentrate on detection method of the video abnormal events which have no concrete class definition.To this end,we divide process of detecting abnormal event into two steps.First we extract feature vectors from the original video.After the feature extraction,we use corresponding abnormal detection model to detect abnormal event.In the feature extraction,we employ independent component analysis,sparse coding and sparse autoencoder.In the abnormal detection model,we detect abnormality based on the reconstruction errors of tensor factorization or prediction errors from recurrent neural network.And based on these two models,we propose two abnormal event detection systems which are based on reconstruction errors and prediction errors.In addition,we analyze the characteristic of the two systems and point out that the two systems detect abnormality based on two different ideas.We then test our systems on a dataset of crowd abnormality.At last,we also analyze the ability of processing dynamic information in the two systems.We find out that the system based on the reconstruction errors embeds the dynamic information in the features during the video feature extraction.As for system based on the prediction errors,base on the analyzation of states of the hidden recurrent layer during the test,we demonstrate its recurrent neural network's ability to learn and process the dynamic information in the video.In conclusion,we design and implement two effective abnormal event detection systems.And base on the test of these systems,we demonstrate their abilities to process dynamic information in the video.