Research On Abnormal Target Detection Technology In Surveillance Video

Detection of abnormalities and abnormalities of the population is becoming more and more important for widely deployed intelligent monitoring systems.Most of the current methods for detecting remnant abnormality rely on the tracking of the carrier.The detection methods of the crowd abnormality are mostly to detect and track the trajectory of the pedestrian target and then analyze whether the behavior is in accordance with the normal state.These methods can lose targets or capture inaccurate targets.To this end,this thesis adopts the tracking-free detection technology mainly to analyze and study the two abnormal targets under the surveillance video.1.Abnormal remnant detection method based on dual feature and CNN combinationThis method aims at removing false target detected as abnormal remnants.It combines both initial detection of stationary target areas and elimination of pseudo-stationary targets(including the varing illumination patterns and pedestrians,etc.).Firstly,the background difference method is used to obtain the foreground target,and then,the initial stationary target is obtained by time domain statistics of the foreground.Two kinds of pseudo-stationary targets caused by illumination are excluded:one is the area similar to the background,which is eliminated by the directional gradient histogram feature;and the other is the light spot caused by the light mutation,which is eliminated by a method of skew distribution.Finally,this thesis uses the CNN network training method to eliminate the pseudo-stationary targets generated by pedestrian detention for the misdetections caused by the static pedestrians.2.Abnormal crowd detection method based on co-occurrence matrix features of the optical flow imageThe crowd panic caused by emergencies is often accompanied by changes in speed and density,and the method is proposed for how to accurately determine the crowd abnormalities caused by panic.It is based on the characteristics of the co-occurrence matrix of the optical flow image.The symbiotic matrix features from each optical flow image are extracted;then these features are clustered by K-Means;finally,the LDA model is trained by the codebook vector formed by all normal video frames to obtain the model parameters and utilized it to determine abnormality of each video frame.Experiments with several public datasets and our video set show that the proposed methods can detect abnormalities and alarms in an effective and timely manner.