Research On Perception Optimization And Content Analysis Algorithms For Intelligent Video Surveillance System

Intelligent Video Surveillance(IVS)system has changed the artificial monitoring screen monitoring and content analysis mode in traditional video surveillance system,to a certain extent,for video surveillance to provide a broader prospect and development space.However,how to obtain valuable object information based on ordinary sensing devices is one of the core problems of IVS.To solve this problem,the perception optimization scheme is studied according to distributed video coding(DVC)and video recovery to obtain perceived videos with higher quality.Then,the content analysis scheme is researched based on video object segmentation algorithm to realize the reliable extraction of objects information from mass perceived videos.The main work of this dissertation is summarized as follows:1)The research on prediction-based video coding optimization algorithm.Aiming at the requirements of low-power and high-efficiency of sensing devices for IVS system which is based on wireless multimedia technology,the DVC scheme is studied.On account of the decoded videos quality reduction problem caused by the interference factors such as videos background environment and objects motion,this dissertation proposes the prediction-based side information(SI)generation algorithm.In this algorithm,the SI prediction model is firstly obtained based on the video blocks attributes and the motion relationship between them.Then,the SI prediction model is employed to generate more accurate motion vectors according to the video blocks attributes,thus improve the SI quality.The experimental results indicate that the proposed method can satisfy the requirements of IVS system which is based on wireless multimedia technology,and alleviate the interference of videos environment and objects motion on SI quality,thereby improving the quality of decoded perceptual data.2)The research on TT rank-based video recovery algorithm.In view of the data quality degradation caused by the influence of light,weather and other factors on ordinary sensing devices,the video recovery algorithm is studied for this problem.Aiming at the video missing problem in the data degradation phenomenon,this dissertation proposes the video completion algorithm on the basis of TT rank.In the proposed scheme,considering imbalanced data distribution and redundant information issues in videos,the approximate tensor generation scheme under the concept of TT-rank is firstly introduced to obtain better-constructed tensor which has balance information distribution and preserves only relatively significant informative data.Then,in order to better map the impact of data in video completion process,the TT-rank-based adaptive weight scheme is utilized to define the weights adaptively in tensor completion problem.Theoretical analysis and simulation experiments exhibit that the proposed algorithm can avoid the replacement of expensive sensing equipments and achieve the optimization of perceived videos.3)The research on AMC-based perceptual video clip object segmentation algorithm.Based on the perceptual video with higher quality,in content analysis,in order to establish an adaptive object segmentation model for massive videos.the video object segmentation algorithm based on absorbing Markov chain(AMC)is firstly designed for perceptual video clips.On account of the negative influence of factors such as relatively violent object motion and larger object deformation,the proposed algorithm realizes accurate object segmentation of perceptual video clips by fully considering the spatiotemporal features,and the segmented results are used as labeled data for perceptual video object segmentation scheme to achieve adaptive object segmentation of massive perceptual data.The introduced scheme firstly constructs the AMC directed graph utilizing the generated object proposals and the proposed weight models.Then,based on the object proposals filtered by absorption time,the re-selection method is introduced to ensure the only object proposal per frame.Finally,the optimization algorithm is employed to generate the segmentations.The experiments verify that the proposed algorithm can segment objects effectively from perceived video clips which have interference factors such as relatively intense object motion and larger object deformation by fully utilizing the spatiotemporal features,and provide the relatively accurate labeled data for perceptual video object segmentation scheme.4).The research on attention-based perceptual video object segmentation algorithm.Based on the proposed perceptual video clip object segmentation method,the attention-based perceptual video object segmentation algorithm is proposed to establish an adaptive object segmentation model for massive video data.Aiming at the negative impact of interference factors such as objects motion and objects deformation on segmented results,the designed scheme is utilized to extract the object information reliably from massive perceptual videos by fully considering the temporal features.The proposed scheme has two branches: deep network model and optical flow prediction model.In deep network model,the attention mechanism is firstly adopted to highlight the object-related features,then,the Conv3 D and the designed attention-based residual convolutional long short term memory(AR-Conv LSTM)are employed to effectively capture the temporal features of video objects.Meanwhile,the optical flow prediction model is designed to lower the negative impact of background motion on segmented results.Finally,the nosing area clean algorithm is introduced to remove the non-target area in segmented frames and obtain the segmentations.The experiments demonstrate that,based on labeled data,the proposed scheme can achieve the adaptive object segmentation of massive perceptual videos which have interference factors such as objects motion and deformation by fully considering the temporal features,and realize the reliable extraction of objects information.