Abnormal Behavior Detections Under Surveillance Video Scenes

The abnormal behavior detection under surveillance video scenes means that the surveillance systems can automatically process the video information without any interference,find the target,mark the abnormal behavior in surveillance videos,and perform real-time alarm.Anomaly detection technology plays an important role in the video surveillance of public areas such as sidewalks and entrances.However,in different monitoring scenarios,due to the influence of some factors such as the change of the viewing angle,the difficulties of defining the types of abnormal behaviors,the change of the video scene,and the change of the non-rigid target,the difficulties of anomaly detection in surveillance video scenes are greatly increased.Under this condition,how to establish an anomaly detection method with high detection rate and adapted to multiple video scenes has attracted extensive attention of researchers in recent years.This article researches and discusses the abnormal behavior detection technology under surveillance video scenes.The main research results include the following aspects:(1)Aiming to take account of itself consistent representation of the moving object and the motion information of the target in the time dimension,the anomaly detection algorithm based on the super-pixels time context is proposed.In the feature representation phase,each frame is processed with super-pixels algorithm,and then whether or not the super-pixel belong to the foreground is judged according to the proportion of the foreground obtained by the Gaussian Mixture model of each super-pixel.Then,according to the gray histogram and position information of the super-pixels,the closest matching super-pixel in the adjacent frame is found.And the super-pixel is represented by the Multi-scale Histogram of Optical Flow feature mean value of the closest matching super-pixels.In the phase of anomaly detection,the super-pixel features of the training set is first learned by sparse combination learning algorithm.In the test stage,determine if the super-pixel is abnormal based on the minimum reconstruction error between the super-pixel feature and dictionary combination set in the test set is determined.(2)Aiming at the problem of the inhomogeneous distribution of normal features and contextual influence existing in the spatial-temporal neighboring features,an anomaly detection algorithm based on prior weights sparse coding and spatio-temporal refinement is proposed.At the feature extraction stage,the dense tracking algorithm is used to extract features.During the training phase,an online dictionary update algorithm was used to model normal features.In the test phase,since the distribution characteristics of the normal features cannot be expressed intuitively using mathematical formulas,the normal feature distribution characteristics are indirectly represented using the degree of importance of the learned dictionary basic vector representing the normal features.After considering the prior weights of the normal feature distribution,the test features are reconstructed.Then the initial reconstruction errors corresponding to the feature features are calculated.Later,considering the temporal continuity of abnormal behavior and normal behavior and the spatial inconsistency of abnormal behavior,a Gaussian refinement of the initial abnormality was performed.Finally,the refined anomaly degree was compared with the abnormality threshold to determine whether it was abnormal.(3)Aiming at the problem that only using convolutional Auto-encoder network could not consider time information,an anomaly detection model based on Bayesian fusion spatial-temporal stream is proposed.Spatial stream model directly reconstruct video frames using a Convolution Auto-Encoder network.Inspired by the LSTM Encoder-Decoder network in the time-series anomalous behavior detections,a convolutional LSTM Encoder-Decoder network is used to reconstruct short-term optical sequence in the temporal stream model.Then,the spatial-temporal stream corresponding reconstruction error is calculated separately.Meanwhile,adaptive thresholds is used for the binary of reconstruction error maps.Finally,the Bayesian fusion method is used to fuse the reconstruction error values to obtain the final fusion reconstruction error value,and to determine the abnormal behavior.