The Research And Application On Human Activity Recognition Based On Key Point Detection,Multi-classifier Fusion And Decision-making

Image-based human activity recognition(HAR)is an intelligent technology that combines computer vision and artificial intelligence(AI)to realize HAR.It has been widely used in special population monitoring,human-computer interaction and other fields,and has become one of the research hotspots of AI.At present,the research of image-based HAR technology is mainly based on a single classifier,which is achieved by optimizing feature information and improving recognition algorithms.Although it has made a lot of research progress,there are still research bottlenecks in the improvement of model versatility and accuracy.A single classifier usually only has a good recognition effect for several kinds of activities that have been trained.After the activity type changes,the parameters need to be re-adjusted for model training,and the generality of the model is poor.At the same time,the classification effect of different classifiers may also have greater uncertainty.Therefore,it is a feasible method to improve the accuracy of image activity recognition by using the complementary characteristic of different classification results.This method can also make reasoning and decision-making for a single uncertain result,which has a certain versatility.This thesis takes image-based HAR and data fusion methods as the research point,and proposes a HAR system based on the combination of bone key point detection optimization and multi-classifier fusion.The specific research contents are as follows:1.In the process of detecting bone key points,there may be missed and wrong detections,resulting in some key points without corresponding position information.This thesis proposes a curve fitting and filling algorithm for bone key points.The proposed algorithm uses a quadratic function to perform curve fitting and filling on the position information of the missing points according to the positions of the key points in the two adjacent frames of the missing points,so that the position information of the missing points is closer to the actual value,which can provide correct data support for the next feature extraction and classification.2.Using the characteristic that the classification results of multiple classifiers are complementary,a multi-classifier fusion human activity recognition system named tf.OpenPose-SNK-fd is proposed based on the D-S evidence theory.The proposed system firstly judges the degree of correlation between the results of different classifiers according to the Pearson correlation coefficient,and then uses the Lance and Williams distance to determine the reliability as the weight of the inconsistent data to redistribute the belief function,and finally uses the D-S combination rule to fuse all data to improve the recognition accuracy.3.Aiming at the smart prison scenario,two kinds of cooperative activities in the specific environment of the prison are determined:handing things and moving things.This thesis collects and builds cooperative activity data set,and using the proposed data fusion algorithm to conduct related research on cooperative activity recognition.In this thesis,the HMDB51 public data set and self-collected data set are used as experimental research data.The experimental results show that the accuracy of activity recognition is improved by about 1.2%,which proves the effectiveness of the proposed algorithm.The research results of this thesis can be applied to various scenarios such as patient monitoring and smart prisons,and have certain practical significance and application value.