Research On Human Activity Prediction Of Ongoing Activities From Videos

Human activity prediction under the background of sparse data has a wide range of application scenarios in real life,not only can it infer the process of events with data missing,help to restore the truth in some cases,but also it can predict consequence of abnormal behavior.The goal of Human activity prediction is inferring the category of the ongoing activities from partial videos.However,there is few research about human activity prediction in the current domestic and abroad,especially in the case of incomplete data.This paper mainly studies the real-time prediction model and optimization features two issues,the main work is as follows:Firstly,to solve the use of limited observational video data to infer the category in real-time with limited calculation of resources,and balancethe prediction accuracy with timeliness,this paper propose SCSW prediction algorithm and sparse coding strategy based on the sliding window.Each activity divided into a plurality of ordered sub-segments,sub-segments of video training extract temporal characteristics as the basis vectors,and apply sparse coding to reconstruct,using of a sliding window strategy,followed by calculation of the test video window and sub-segments training video corresponding to the similarity between the sub-segments.Finally,the union of all the similarity test video to get the posterior probability of belonging to a certain class.Secondly,in order to achieve better prediction accuracy,it need to capture the distinct features which can maximize the difference between the classes and minimize inter-class differences,this paper proposed a feature optimization algorithm based on statistical properties.Expanding the feature samples of the same segments of the same activity,calculating frequency and variance of each feature in the sample,after several iterations,achieving some stable and representativefeatures of each activity of each segment.Furthermore.based on the proposed feature optimization algorithm,this paper using neural network algoritlm to predict the activity with videos missing the front portion.Finally,two public interactive activity recognition databases are used in the experiments.Results show that the proposed SCSW method can improve the real-time performance of activity prediction at the same time remain at high prediction accuracy,as well as verify the validity of the proposed feature optimization algorithm.