| Video processing technology has always been the focus of research in the field of computer vision and pattern recognition.In recent years,object behavior detection in video-related tasks has attracted extensive attention from researchers due to its high application value.The object behavior detection task requires object positioning and behavior classification.The current common method is to use a deep neural network to process the object behavior detection task,thus proposing two frameworks of two-stage detection and one-stage detection.Among them,the two-stage detection is to locate the object first and then complete the behavior classification.Many excellent two-stage works use the technology of relational modeling to increase the mining of video information,thereby improving the effect of object behavior detection.However,current relational modeling methods all rely on object detectors that are independent of the network,making it difficult to effectively apply existing relational modeling methods to a one-stage framework that does not use object detectors and simultaneously completes object location and behavior classification.In view of the above problems,the thesis proposes a one-stage object behavior detection framework based on relational modeling,which uses the teacher network to complete relational modeling and guide student network to learn relational modeling knowledge,and effectively integrate the advantages of fast speed,end-to-end of one-stage detection framework and good effect of using relational modeling method.At the same time,the thesis studies the effect of different teacher-student network training methods on relational modeling knowledge transfer.The main work of this thesis is as follows:(1)A one-stage object behavior detection framework based on relational modeling is proposed.The framework uses a teacher-student network and is a fully end-to-end onestage detection framework.In the training phase,the teacher network and the student network are trained from scratch,and the teacher network uses the label position information prior as an additional input to complete the relationship modeling and guide the student network to learn the relationship modeling process by optimizing the hint loss.In the inference stage,only the student network is retained to complete the behavior detection task independently.(2)A teacher-student network learning method based on dynamic mutual learning is proposed.Aiming at the problem of how to better train the teacher-student network to improve the effect of knowledge transfer in relational modeling,the thesis proposes and studies a strongly coupled learning method based on shared convolution,a one-way learning method based on gradient blocking,and a mutual learning method based on hint loss,and finally proposes a dynamic mutual learning method with adaptive ability.Based on the one-stage object behavior detection framework based on relational modeling proposed in this thesis,the effect is verified.(3)An object behavior detection system based on B/S architecture is designed and implemented.The system design,interface design and implementation are carried out.Users can upload videos on the webpage,call the object behavior detection algorithm model deployed on the server,detect the object behavior and get the processing results. |
