| Abnormal video behavior detection refers to the use of computer vision technology to automatically analyze and extract abnormal events in video sequences.With the increasing demand for anomaly detection in outdoor daily activities and home monitoring,abnormal video behavior detection has important research significance.Due to the low frequency and variety of abnormal events in daily life,data containing all abnormal samples are difficult to obtain,so existing methods usually learn the characteristics of normal samples in a self-supervised way.In this thesis,a selfsupervised learning approach is used to model normal patterns,and then detect abnormal events different from normal patterns during the testing phase.The main contents of this thesis include:Aiming at the blindness problem of pattern mining in self-supervised learning,this thesis first proposes a video anomaly detection method based on an autoencoder and self-attention mechanism.This method embeds the Transformer module based on the self-attention mechanism in the U-Net network to fuse the local detail information and global spatial information learned by the two networks,and can effectively locate abnormal behaviors in the video.In the training phase,this thesis introduces dynamic image constraint to guide the network to pay attention to abnormalities in motion areas,thereby improving the accuracy of abnormal behavior detection.Experimental results show that the proposed algorithm can effectively detect abnormal behavior in outdoor and indoor videos.In order to alleviate the problem of weak robustness caused by unlabeled supervision in self-supervised learning,this thesis then proposes a mutual supervised anomaly detection algorithm based on generation consistency,which combines normal patterns and pseudo-label information for anomaly detection.The algorithm simultaneously mines normal patterns and hidden label information of different input lengths in a mutual learning way.This method provides an enhanced regularity to the model by introducing unified label information that is not related to the video,thereby improving the robustness of the model.In addition,in order to meet the needs of improving model performance in resource-constrained scenarios,an auxiliary network with dense inputs is also constructed in this thesis,and the feature learning ability of low-resolution sparse input networks is improved by mutual supervision learning.Experiments show that this method can improve the accuracy of anomaly detection in low-resolution images.The real-time test speed reaches 80 fps,which shows that the method can detect abnormal actions in the video at real-time speed.For indoor fine-grained behavior anomaly detection,this thesis designs a difference attention module based on the previous two works to improve the attention to detail errors.This method proposes a cross-attention network C2Net based on video content-independent information,which embeds video information into contentindependent input for anomaly detection.C2Net includes a prediction model to model video content and a reconstruction model to learn content-independent information.In addition,a cross-attention module is also designed to separate the two modalities,thereby improving feature expressiveness and anomaly analysis performance.During the training phase,a homogeneous auxiliary network with sparse inputs helps the feature learning of the backbone network through mutual learning.Experimental results on indoor fine-grained motion datasets demonstrate the effectiveness of our approach in video anomaly detection.For the limitation of model mining in self-supervised learning,this thesis conducts research work on model design,learning method,pseudo-label design and attention mechanism.Experimental results on multiple datasets show that the algorithm can effectively detect abnormal behavior in video. |
