Analysis Of Abnormal Behavior Of Weak Supervision Based On Multi-Scale Time Modeling

Abnormal behavior analysis means identifying abnormal behavior from video sequences accurately and quickly.It is one of the most important research contents of computer vision.Abnormal behavior analysis has a wide range of applications,including motion-acquired action recognition,video surveillance,video retrieval,and public safety analysis.In recent years,incidents which affect public safety happens a lot,such as terrorist attacks and stampede.How to find and propose efficient solutions to unexpected abnormal behaviors and events quickly,and how to automatically analyze massive video in order to replace 24-hour real-time monitoring staff are hot issues that cause widespread concern in the field of public security at home and abroad.So how can we find emergencies in time while occupying the least human and material resources?To handle this problem,the abnormal behavior analysis system comes into being and stands out in the field of computer behavior recognition.With the efforts of scholars at home and abroad,many research results have been achieved in the field of abnormal behavior analysis in video monitoring system.At present,there have been studies on behavior modeling and classification by using human trajectories tracking,moving object detection and so on.However,the traditional abnormal behavior analysis has certain disadvantages.When the algorithm is applied to massive surveillance video data,there are problems of poor recognition effect,slow processing speed and algorithm redundancy,which not only causes a lot of waste of resources but also affects the overall analysis efficiency of the algorithm.To handle this problem,this thesis make thorough research on improving the recognition accuracy of abnormal behavior analysis and reducing resource occupation.In order to solve the problem of abnormal behavior recognition in video surveillance image,the network structure is optimized based on the deep convolution network and the loss function is improved,thus the accuracy of abnormal behavior recognition is effectively improved.The major work and research results obtained are summarized as follows:1.Based on the behavior time series characteristics of multi time dimension C3D network,the model is established.By studying the feature extraction method of the convolutional neural network,we analyze the 3D convolutional neural network in detail and improve the C3D network.On this basis,the time convolution kernel is extended to construct a multidimensional 3D convolution network which can capture the time characteristics of detailed behaviors.By using the multi-dimensional convolution kernel in the time domain,we can extract the features of various time ranges from complex actions,thus increasing the diversity of behavior feature extraction.2.Based on multi-instance regression loss,weakly supervised abnormal behavior is recognized.Because the boundary between normal behavior and abnormal behavior is usually ambiguous,unlike common event classification methods,abnormal classification is transformed into a regression problem.By defining the loss function as an orderly regression loss of the predicted probability of the abnormal event and the normal event,the abnormal score of the abnormal event must be higher than the abnormal score of the normal event,and vice versa.By constraining the regression ranking of the scores of the abnormal and normal behaviors,the algorithm proposed in this thesis can increase the discrimination between normal and abnormal events,and thus identify the abnormal events from the video events more accurately.Because the training process uses a weakly supervised learning framework,only video-level tags are used and accurate time tags and classification tags for abnormal events are not needed.Therefore,it effectively reduces the requirements of the algorithm for data annotation,makes the algorithm have stronger expansibility and scene adaptability,and has certain practical application value.3.Finally,in order to verify the effectiveness of the weakly supervised abnormal behavior recognition network based on multi-dimensional time modeling in richer feature learning and recognition accuracy improvement,experiments were performed on the UCF-Crime public behavior data set.The experimental results show that the weakly supervised abnormal behavior recognition algorithm based on the multi-dimensional time module can effectively extract the features of different time dimensions and improve the accuracy of network recognition.Compared with other methods,the effect on the abnormal behavior recognition task is improved by 1.5%.This thesis proposes an abnormal behavior analysis algorithm based on deep learning to optimize the deep learning network model feature learning method,the overall network framework and recognition accuracy.It can also accelerate the recognition process and reduce the resource occupation while ensuring the recognition accuracy,thus complete the abnormal behavior identification tasks and time positioning tasks more accurately.