Research On Abnormal Behaviour Detection And Frame Rate Up-conversion In Video Monitoring System

Under the background of social security problems and high difficulty in social management,the research of video monitoring system is particularly important.This paper focuses on abnormal behaviour detection and frame rate up-conversion in video monitoring system.In the research of abnormal behaviour detection,we propose a method of abnormal behaviour detection based on Dense Trajectories and Fisher Vector to improve accuracy and another method of abnormal behaviour detection based on MPEG flow video descriptor to optimize detection time.In the research of frame rate up-conversion,we improve the block-based motion compensation interpolation algorithm and propose a method of frame rate up-conversion based on bidirectional motion estimation and median filtering.Next,we will give a brief description of these three methods proposed in this paper.1.Abnormal Behaviour Detection Based on Dense Trajectories and Fisher Vector: In order to extract more representative feature of video and improve detection accuracy,this paper proposes a method of abnormal behaviour detection based on Dense Trajectories(DT)and Fisher Vector.In feature extraction,DT feature extraction method is used to densely sample video,track feature points and describe trajectory shape,appearance,motion and motion boundary of the sampled feature points.In feature coding,Fisher Vector method is applied to compute the high-order statistic,thus converting low-level features into high-level video representation.Finally,experimental results show that compared with other methods of abnormal behaviour detection,the accuracy of our proposed has been improved,and it is effective both in non-crowd and crowd scenarios.2.Abnormal Behaviour Detection Based on MPEG Flow Video Descriptor:In order to reduce the computational complexity of feature extraction and optimize detection time,this paper proposes a method of abnormal behaviour detection based on MPEG flow video descriptor(MF).In feature extraction,MF feature is used to represent the appearance,motion and motion boundary of video and MPEG flow is employed to replace Optical flow,which greatly reduces the computational complexity and improves the processing rate of video.In feature coding,sparse coding algorithm is applied to reconstruct sparse representation of the original video,thus converting low-level features into high-level video representation.Finally,experimental results show that our proposed is more efficient than other methods in terms of accuracy and detection time,and it is effective both in non-crowd and crowd scenarios.3.Frame Rate Up-Conversion Based on Bidirectional Motion Estimation and Median Filtering: In order to reduce the computational complexity of motion estimation without affecting the final accuracy,we proposed a method of frame rate upconversion based on bidirectional motion estimation and median filtering.In motion estimation,we design a two-stage method based on bidirectional motion estimation,which consists of coarse searching and fine searching.Coarse searching unit is responsible for calculating the initial motion vectors,while the fine searching unit is responsible for fine-tuning the motion vectors adjusted by error.In motion refinement,we propose an algorithm of Angular-Distance Median Filter(ADMF)to correct wrong motion vectors.Finally,experimental results validate the effectiveness of our proposed in terms of PSNR and SSIM.