Research On Video Sequence Detection Algorithm For Health Management

With the aging of the population and the improvement of living standards,people begin to pay more attention to their own life and health.In this social context,the study of video analysis algorithms in the field of health has gradually come into the field of researchers' vision,and has been highly valued,and has gradually become a research hotspot.The research of video analysis in the field of health mainly includes disease diagnosis based on medical image and health monitoring based on surveillance video.In the field of disease diagnosis,most of the current method is based on medical image data,which is generally judged by doctors according to their own experience and professional knowledge,but this generally has a certain rate of misdiagnosis.In the field of health monitoring based on video data,people's behavior is mainly analyzed to determine whether they are in a dangerous state.However,purely manual detection is costly and inefficient.Therefore,it is of great significance to carry out health monitoring research.Compared with traditional video analysis algorithms,health-oriented video analysis algorithms have certain particularities,which are manifested in the following aspects:(1)data level.There are relatively few video datasets in the field of health.In addition,there are some problems such as small number of samples,unbalanced distribution of samples and too large intra-class spacing.(2)Difficulty in extracting time boundary.It is difficult to accurately extract the time boundary of video content because it cannot make full use of the implied time sequence information in video data.Video content classification is more dependent on the location of time boundary,and the unsatisfactory location of time boundary will have a negative impact on video content classification.(3)Low efficiency.The time span of video data is large,the phenomenon of data redundancy is serious,the processing process is time-consuming,and the feedback cannot be given in time.To solve the above problems,two health-oriented video analysis algorithms are proposed in this thesis:(1)This thesis proposed a video analysis method based on DE-Res Net framework.This algorithm is designed to address the data-level challenges of video analytics in the health domain.Data sets in the health field generally have problems such as insufficient data,large sample spacing within the class,and unbalanced sample distribution.For this reason,a hybrid data enhancement method is proposed in this thesis,and a DERes Net algorithm framework is proposed based on residual neural network.Moreover,this thesis also applied this algorithm to the task of gastrointestinal disease detection.Experiments show that DE-Res Net algorithm has significant advantages in dealing with small-scale data sets with unbalanced sample distribution and large intra-class spacing.(2)This thesis proposed a video analysis method based on C3D-BMN framework.C3D-BMN algorithm mainly includes three parts: feature extraction module,time boundary extraction module and classification module.The C3 D network feature extraction module is designed to fully extract the spatial and temporal features of video data.BMN network,as the time boundary extraction module,aims to extract the time boundary of video content based on the spatiotemporal characteristics of video.The function of the classification module is to recognize the video content based on the time boundary and the spatio-temporal characteristics of the video.In addition,the C3 DBMN algorithm is also applied to human behavior recognition task to prove its effectiveness.The experimental results show that the multi-mode fusion technology can further improve the detection effect of the model.