Research On Key Issues Of Human Action Recognition Based On Noisy Spatial Location Information

Human action recognition aims to describe human actions, as well as to predict following actions, by analyzing human motion data captured by various sources, and has valuable theoretical significance and broad application prospects.The main work of this paper includes two key issues existing in current human action recognition studies with noisy spatial location information:optimizing spatial location information and recognizing human actions.Regarding the location information obtained under noisy environments, we make use of geometric constraints among the sides of a polygon which is formed by distance measurements between neighboring sensor nodes, adopt the Cayley-Menger determinant to establish a statistical hypothesis testing model, and then identify NLOS channels that damage distance measurements. Extensive simulations are carried out and show that the proposed method is able to successfully identify over 80% NLOS channels.In order to recognize human actions, location information of a human body is obtained under various motions by using the UWB based Ubisense realtime location system, features (including motion features, frequency features and statistical features) are extracted from the location information based on machine learning theory to form time-series feature vectors that describe different actions, and finally, BP is adopted to train the human action recognition model. Experimental results illustrate that the proposed model is able to identify six different daily actions with the overall successful rate above 82%.In summary, human action recognition based on noisy spatial location information is tackled in this thesis, and the proposed method not only improves the positioning accuracy and balances the recognition accuracy with computational costs, but also demonstrates advantages in privacy protection, convenience for wearing and maintenance.