Research On Visual Enhancement And Action Analysis For Surveillance Video

Since the 21st century,benefited from the development of imaging,photoelectric,multimedia and network technology,video surveillance system has become increas-ingly popular in our daily life,and has been playing a more and more important role in security,traffic management,automatic control,medical care,education and many other fields.With the vigorous promotion of major national security projects and Smart City construction,nearly 200 million professional surveillance cameras have been de-ployed in China.Video surveillance is now a super-huge market of over 100 billion in this country.The successful application of video surveillance can not be achieved with-out the booming development of computer vision technology.Computer vision studies how to use computers to process and analyze images(including pictures and videos).In practical applications,the number and density of surveillance cameras deployed world-wide has been soaring rapidly.Thus,on the one hand,people have put forward higher requirements for the quality of surveillance video,which can be achieved with help from image processing technology;on the other hand,in order to handle massive amounts of video data,the surveillance system is also moving from relying solely on manual recognition to computer-assisted management based on video analysis technology,and is hoped to eventually achieve fully automatic intelligent control.In this process,im-proving the understanding of surveillance video content by humans and computers has always been a top priority in the development of video surveillance system.For human observers,higher resolution,sharper structural edges,and more complete context infor-mation can all contribute to a better understanding of the video content.For computers,how to abstract planar lattice data into meaningful information,how to analyze the ab-stract information,and how to process high-density data like video stream in real time is the key to getting a computer understand what is happening in video like a human being.This thesis revolves around improving humans and computers' understanding of surveillance video content.From both the bottom layer image processing and high-level video analysis,this thesis studies the visual enhancement of dim-light surveillance video and the real-time detection of specified behavior in surveillance video stream.Specifically,the main contribution and novelty of this thesis can be summarized as follows:1.In the visual enhancement of dim-light surveillance video,this thesis proposes a method to enhance the nighttime surveillance video through gradient domain op-erations.Daytime video data taken by the same camera can be used to acquire a background image.The nighttime surveillance image is processed by the adaptive gradient enhancement algorithm proposed in this thesis.After that,it is merged with the daytime background image in gradient domain and reconstructed.As a supple-ment,motion detection technology is used to get the moving foreground of night-time video,which is separately enhanced to avoid loss of information.Extensive experiments on real outdoor nighttime surveillance video show that the dim light surveillance video enhancement algorithm in this thesis can significantly improves the visual quality of nighttime surveillance video and effectively suppress the arti-facts such as foreground blur and visual discontinuity in the existing methods.2.In the real-time detection of specified behavior in surveillance video stream,this thesis proposes a multi-feature cascade detection architecture using simple to com-plex features.Easy-to-differentiate video clips are filtered out using simple features with low computational complexity.And features with higher computational com-plexity are extracted from indistinguishable clips for further judgment.This kind of cascading architecture can well balance performance and efficiency.It is suitable for processing data with sparsely distributed positive samples,and more samples can be used to improve model performance during training.In cooperation with the hospital,we collected a large number of actual surveillance video in the wards and manually annotated the behavior to be detected.On this data set,the multi-feature cascade detection architecture of this thesis achieves a positive recall rate of 97.68%while maintaining a false alarm rate of 0.172%.And it runs 42.7fps on average on motion-intensive video stream.The average processing speed meets the require-ments of practical applications.