The Research On Method Of Video Event Structured Description

With the development of video event description,text-based description and content-based description are no longer suitable for describing events that occur in traffic surveillance due to their shortcomings.The research content of this thesis is the description and recognition of the event in the traffic video,which is proposed for the unstructured of the video,and aims to provide the hierarchical expression of the video content.The goal is to organize the information that can be understood by the computer and people in terms of semantic relations,such as deep learning,ontology,and logic description language,for traffic monitoring video content.The specific research contents are as follows:Firstly,a new rich semantic model,namely,video event structured description model(VSD),is designed for the unstructured characteristics of video content.The model extracts information such as objects,attributes and relationships in the video,and forms the video ontology,which can effectively describe the video content to reduce the gap between low level semantics(feature)and high level semantics(event).For the actual creation process of traffic video ontology,two methods of constructing using Protege software and converting from database are proposed,and the conversion process from relational database to domain ontology is analyzed in detail.Experiments With the help of the Protege tool,the implementation of the structured description model of traffic video events was carried out to provide examples for the following event description studies.Finally,the descriptive model is used to score the relevant rules,and the superiority of the proposed model is displayed intuitively.Secondly,in the aspect of object description,we study the two fields of pattern recognition and description logic,that is,object semantic representation from traffic object recognition detection and object relation mapping reasoning.In the process of video object recognition detection,the structural characteristics of convolution neural network are explored,and a network structure suitable for traffic object recognition is proposed.Parameter optimization is carried out.Combined with R-CNN object detection method,(Including the RCC space topological relation and the shape of the space shape)in the structured description of the video event,and use the ontology logic language to carry out the object.In the study of the object relation mapping,we give the concrete definition of the object spatial relation in the structure description of the video event Relationship description.Finally,an object consistency reasoning algorithm is introduced to further deduce the semantic relation between traffic objects.Finally,an event description method based on scene context is proposed by combining the video event structure description model and object mapping.That is,the space,time,and special value context in the traffic video are preliminarily set,and the semantic information of each frame in the video is extracted by using the relationship with the target object.Based on this,the concept of event unit is introduced in traffic event identification,and an event recognition method based on Bayesian classifier and finite state automaton is proposed to realize the recognition of simple and complex traffic events.The feasibility of the proposed event description method and identification method is proved in the experiment.