Abnormal Event Detection And Object Recognition In Surveillance Videos Based On Deep Learning Method

In recent years,artificial intelligence(AI)begins to penetrate into all areas of life along with the rise of the concept of Internet of Things and the development of deep learning.The security industry,as an important means of maintaining social stability and people's personal safety,has been greatly developed due to the industrial empow-erment of AI.In order to continue to promote the development of the security industry,government has proposed the concept of the "safe city".The foundation of building the" safe city"is information perception,and building a comprehensive security monitoring network is an indispensable means.Therefore,video surveillance technology has been more and more widely used,and a large number of monitoring devices are deployed in public places such as campuses,streets,and communities,resulting in massive mon-itoring video data.However,most of the traditional video surveillance technologies only provide acquisition,storage and review functions,and there are many drawbacks.Nowadays,with the development of AI,how to intelligently analyze and process mas-sive video data,extract useful information from it,and identify the target in the video has become a hot research topic in the field of AI.This paper focuses on the task of abnormal event detection,face recognition with different resolutions and person re-identification in intelligent video analysis,and the research work mainly includes:(1)detecting abnormal events in surveillance videos,such as fighting,chasing,and crowd gathering;(2)identifying abnormal or suspicious pedestrians in surveillance videos.Deep learning has made great progress in the fields of image recognition and speech recognition,which provides an opportunity for intelligent processing of surveil-lance video.This paper uses deep learning tools to carry out research work and main innovations are as follows:1.For the detection of abnormal events in surveillance videos,a feature recon-struction method based on adaptive multiple auto-encoders is proposed.This method is aimed at the problem of uneven distribution of positive and negative samples in the detection of abnormal events.Only normal events in surveillance videos are used as the training data,the spatial and temporal information are integrated to train the auto-encoders.The learned auto-encoders can be used to fit the spatio-temporal distribution of normal events and achieve the detection and localization of abnormal events.Since this method only requires normal events as training data,the complexity of the task is greatly reduced.Experiments show that this method can accurately detect and locate abnormal events in surveillance videos.2.In order to further improve the accuracy of abnormal events detection,a feature extraction and anomaly detection method using principal component analysis network is proposed.The method uses the unsupervised deep learning model PCANet to ex-tract the high-level feature representation of normal events,and then uses the clustering method to cluster the extracted high-level features to obtain the clustering center of nor-mal events.For test data,the distance from the cluster center is used to determine its outliers,thus detecting abnormal events.3.Aiming at the problem that the recognition accuracy of the face image in surveil-lance videos is low due to different resolution,a face recognition scheme between high and low resolution using convolutional neural networks is proposed.Two methods are proposed in this paper:(1)Face images of different resolutions are normalized to the same size using linear interpolation,and then the deep convolution model is trained using the mixed samples;(2)Improving the architecture of deep network.Regression loss is used in model training.Specifically,the low-resolution face image and the high-resolution face image with corresponding identity are pairing as the model input,the corresponding high-resolution face image and identity information are used as super-vision information.The distance behtween high and low resolution face images in fea-ture space is reduced by optimizing classification and regression losses simultaneously,thereby achieving an improvement in recognition accuracy.4.Aiming at the problem of low recognition accuracy caused by factors such as human body occlusion and body misalignment in the person re-identification task in surveillance videos,a solution using deep learning is proposed.Two methods are proposed in this paper:(1)A deep network combining classification model and multi-scale matching model is proposed.The classification model is used to obtain the high-level feature maps.On this basis,the multi-scale matching model is used to convolve the feature map at different scales,so that the semantical correlation can be captured to overcome the effects of body misalignment;(2)The second-order pooling is used to realize the attention pooling,thereby effectively extracting the features of the human body region,overcoming the interference of external factors such as background,so as to improving the accuracy of person re-identification.