Research On Abnormal Behavior Detection Of Indoor Person

Intelligent monitoring and intelligent security are quite significant in the construction of intelligent buildings.Abnormal behavior detection algorithms based on Computer vision can effectively improve Security Guard's efficiency,prevent external risks,and improve the automation levels of building's monitoring systems.Although many abnormal behavior detection algorithms based on surveillance video have been well developed,they still have problems in real-time performance,detection for real-world scenarios.Based on deep learning methods,we have proposed three efficient abnormal behavior detection networks.The first method is based on multi-instance learning which is a supervised method.This method relies on a training set with both normal and abnormal behavior samples,and is constructed based on a two-stream structure with optical flow stacks and video clips as input.The collaborative spatiotemporal convolution with residual structure has been employed for feature extraction and a time information network composed of GRUs was built as classifiers.Moreover,the concept of multi-instance learning was introduced into the training producer,i.e.,a long video is regarded as multi-packages,each clip is regarded as an instance,hinge based ranking loss between anomaly scores of instances are constructed.During its inferring stage,the anomaly score of each video clip was regarded as output.Second network is a spatio-temporal autoencoder based on convolution-deconvolution,in which Conv GRU layers were introduced for coding-decoding along the time dimension.Different from the first method,only normal data was used in the training stage.Considering the fact that one there are abnormal events,the reconstructed clips will have larger errors,so evaluating the reconstruction errors can perform fast abnormal behavior detection.Thus,this method realized abnormal behavior detection with light computational consumption.Following the normal future frame prediction style,we then assume that the abnormal future frames or sequences are unpredictable,and proposed the third network based on two major branches: prediction and reconstruction.The prediction branch is a generative adversarial network which trained with loss functions including appearance,gradient,time consistency to ensure the quality of generation,we have also employed the loss function of least squares GAN to ensure training stability.The reconstruction branch is an autoencoder that adds the latent space memory enhancement constraint based on the attention mechanism.based on the attention mechanism.The weighted frame-level pixel error between the real frame and the generated frame produced by two branches was calculated through the regional sliding windows.Normalizing this error to [0,1],and setting a proper threshold,then the abnormal events can be detected.Those three methods have achieved the AUC indicator of 0.823,0.739,and 0.861 on CUHK Avenue Dataset respectively,and have also achieved comparable results on other datasets.We have also developed a software system combined with the proposed models for abnormal behavior detection in real-monitoring scenes.The system was run on a Jetson Agx Xavier board and with a multi-process structure.The models are accelerated with Nvidia's Tensorrt technology.The well-designed system performs a frame of 256×256resolution only 18.9ms,quicker 4?5 times than most of the CPU-based detection systems,satisfing the real-time requirement.