Motion Analysis And Recognition Based On Inertial Sensors

Human motion analysis and recognition based on inertial sensors has becoming anemerging field in pattern recognition field. It has higher operability and practicality byovercoming shortcomings and limitations of the traditional video-based motion recognition.Its essence is collecting motion data of humans through fixed inertial sensors, and the data istransformed to the PC via a wireless module. Then, the data preprocessing, feature extractionand action classification are carried out. A motion capture system is based on the inertialsensors including accelerometers, magnetometer, gyroscope, and these inertial sensors makeup sensor nodes communicating through wireless module. This paper presents the method ofcollecting human motion data by use of inertial sensor-based motion capture system.The paper first analyses related parameters of human motion and their measurementmethods from the perspective of sports biomechanics, and confirms the acquisition programbased on inertial sensors. On this basis, the paper analyses the human joints, links, jointangles and attitude angles, and explicit body structure and multi-rigid-body model with andsimple improvements which identified11sites as sensor wearing position.After a deep analysis of human motion, the paper proposes a novel method of humanmotion analysis named PAMS which divides human motion analysis into three levels of pose,action and motion. The essence of the PAMS method is that a series of poses constituteactions and a series of actions constitute motions.The paper then sums up31kinds of actions in the human motion, and other complexactions and motions are a combination of these basic actions. Using motion capture systembased on inertial sensors to collect human action data which are all stored in a new databaseto establish a human action database to prepare for the post-study and the usage by otherresearchers.Finally, the action recognition algorithms are studied in this paper. Taking into accountthe system noise, the raw data is firstly pre-processed, and characteristics are extracted fromthe time domain and frequency domain. And the principal component analysis is used toreduce the dimensions of characteristics vectors. The paper then executes an actionrecognition experiment with the method of support vector machine which shows goodclassification.