Research And Implementation Of Semantic Modeling For Surveillance Video

Nowadays,the surveillance video is widely used in public places such as airports,subways,stations,and campus,which goes deep into people’s lives in all aspects.With the development of the market size of video surveillance,these massive surveillance videos pose new challenges to video surveillance systems.However,due to the large number of surveillance videos and unstructured features,the use of surveillance video has not been effectively utilized currently.In order to effectively use surveillance video data,this article focuses on video structuring technology.The main research contents are as follows:(1)A video foreground extraction algorithm based on improved ViBe is proposed.Since most video cameras of the surveillance video are fixed,the change in the background of the surveillance video is small,and the part of interest to the surveillance video is also a foreground part of the video.Therefore,we need to extract the foreground of the video which is used as a basis for follow-up research.The Vi Be algorithm is fast and simple to implement,but there is a ghost phenomenon in the foreground which extracted via ViBe algorithm,and its fixed threshold and update probability also restrict the adaptability in the complex scene.This paper proposes an improvement program according to these problems.First of all,we use multi-frames to initialize,and then a dynamic matching threshold is used to decide whether the pixel is a foreground pixel.Then we detect the ghost with frame-difference method,and then update our model and propagate our result to its neighborhood with the dynamic probability.Finally,it is verified through experiments that this improved scheme has a better evaluation.(2)A model combined with attention mechanism based on injected semantic attribute is proposed.At present,the attention mechanism is widely used in the task of generating video description statements.Injecting semantic attributes into the semantic description generation model has also been proved to improve the results.This paper combines the advantages of these two schemes,and proposes a model combined with attention mechanism based on injected semantic attribute.We use the temporal attention to focuses on various parts of the video gradually,and as for semantic attributes,we use the semantic attention to gradually focus on the semantic attributes,and then we dynamically adjusts the utilization of visual information and contextual information through adaptive attention.Finally,it is verified through experiments that this scheme has better effect in the tasks of generating video description.(3)A surveillance video oriented semantic modeling system is designed and developed.we first design the overall architecture of the system,and then describe the data management module,core function module and interaction and display module in the system in detail.The video foreground extraction algorithm based on improved ViBe and the model combined with attention mechanism based on injected semantic attribute are integrated into related modules.Finally,the system was demonstrated.