Research On Collective Motion Analysis And Recognition

Identification and analysis of collective motion are frontier field of pattern recognition and computer vision,provide effective technical means of video surveillance for public places,real-time battlefield analysis and video retrieval,etc.With the rapid development of computer vision technology,target detection and recognition technologies based on image become more and more mature.However,techniques of video still need to further improvement,especially for the behavior recognition.Identification of collective motion is a kind of behavior recognition,and the scenes of collective motion are more complex than the general behaviors.So far,different collective motions are difficult to complete corresponding analysis and recognition task with a fixed algorithm.Compare to single target or multiple targets,the collective motion is an entirety behavior related to the context of targets.And the collective motion character as a large number of targets,complex environment,different velocity of individuals,high density and multi-target occlusion etc.At present,few of scholar research on this subject.Although some achievements have been made in some large scale public datasets,their entry points are different.Therefore,the overall research is still in a relatively dispersed phase.This thesis mainly pay attention on two categories of collective motion.The first is sparse collective motion(e.g.carrier vessel battle group,CVBG),and the second is dense collective motion(e.g.crowded).In this thesis,sparse collective motion is studied by taking CVBG as an example,and dense collective motion is studied by taking crowd as an example.To characterize collective motion,the following research is carried out in this thesis:Firstly,this thesis simulates CVBG sailing videos under surveillance of satellite.The CVBG's datum are very valuable,simulation of the CVBG sailing is significant.Although the satellite technology is difficult to support a wide range of video capture,analog video can verify the recognition algorithm at any time.In order to simulate CVBG's sailing as much as possible,this thesis analyzes the feasibility of radar satellite and optical satellite for CVBG surveillance.The trajectories of warship are cubic Hermit interpolation functions,which can not only guarantee the second derivative but also can control distances among warships to prevent collision.In order to enhance the authenticity of analog video,the aircraft carrier and warship templates come from the Google earth.Secondly,formation recognition and behavior analysis are based on that ship targets have detected.In this thesis,an invariant descriptor,named multi-viewpoint context(MVC),is formed by calculating context in viewpoints selected from the Archimedes spiral.A probability density function(PDF)model,constructed by Laplace transform,integrates local and global information and enhances MVC discrimination.The dimensionality of MVC is not related to the number of CVBG.Moreover,MVC is sensitive to the central warships of CVBG,which is consistent with factual situations.The behavior recognition method is based on Hidden Markov model and verified in different analog videos.For crowd detection,this thesis presents a new local descriptor which is named multiple viewpoint histogram(MVH).This descriptor,based on the radial gradient transform(RGT)of multiple observation points,is robust to scale and pose.The method can deal with different scales of the crowd image block without image scaling.EdgeBox,an object proposal method,can use a few proposals that cover most of targets to detect crowd without Gauss Pyramid.Sufficient experiments demonstrate effectiveness of the proposed method.Finally,a multi-babel classifier is proposed for crowded scenes based on the dependency among categories.Crowded scenes characterize as many categories,complex relationship,and traditional classifiers perform with inefficiency.In thesis,a spatial-temporal video descriptor which is constructed by CNN and Fisher Vector(FV)encoding can efficiently deal with multi-instance problems.The methods of this thesis are demonstrated by sufficient experiments.The experimental results show that proposed methods outperform the-state-of-the-art ones.Most of data set are actual video or images,and the proposed methods have a significant practical value.Our method can be regarded as some fundamental techniques that show potentials to other related fields,such as event detection or action recognition.