Research Of Moving Object Behaviour Recognition Method Based On Multi-view Videos

With the development of Computer Vision,intelligent video surveillance technology has attracted widespread attention on many real-world applications.Plenty of video data are produced in public palces and important locations,thanks to the multi-view cameras.The main problem of intelligent video surveillance technology is detecting,tracking and classifying the interesting target in the dynamic scene.Then,it analyzes and understands the semantic information of interesting target or scene event.However,many researches on intelligent video surveillance technology are based on monocular videos,which cannot fully characterize the property of behaviors.In this paper,for the problem of safety protection and crowd management,we use multi-view low-rank dictionary learning method to integrate information from multiple videos,such as parking behaviour classification or crowd behaviour classification.The contributions of this thesis are summarized as follows.(1)Parking behaviour analysis and recognition based multi-view videosFor muli-view parking video data,different from analysis method based on one frame,we use the continuous frame sequence of multi-view videos as the state inspection data of roadside parking spaces and adopt KLT(Kanade-Lucas-Tomasi)motion corner detection and real-time compression tracking method to extract the trajectory of the whole process of vehicle parking in the video sequence.Then,we build the vehicle status characteristic descriptor for representation and recognition of vehicle parking behavior.The experiment results on the real-world video data show that the proposed method is effective and can be applied to the roadside parking recognition.(2)Crowd behaviour analysis and recognition based multi-view videosFor the multi-view crowd monitoring data,we present a new method of crowd behavior analysis based on the fusion of multi-view information.In this method,we firstly track the trajectories of crowd behavior in the video and use the correlation of the trajectories to cluster the trajectories.Then,we can extract the characteristic descriptor of crowd behavior and finally achieve the recognition of the crowd behavior captured from multi-view videos.Experiments on multi-view videos data of actual monitoring scene show that our method has good performance on crowd behavior recognition.