Towards Multi-Classification of Human Motions Using Micro IMU and SVM Training Process

Guang Yi Shi; Yue Xian Zou; Wen J. Li; Yu Feng Jin; Pei Guan

doi:10.4028/www.scientific.net/AMR.60-61.189

Paper Titles

DNA Biosensors Based on Self-Assembled Multi-Walled Carbon Nanotubes
p.170

Fabrication and Testing of Hard X-Ray Hourglass Lenses
p.175

Study of Control System for a Micro Programmable Blazed Grating
p.180

High-Surface-Quality and Wider-Modulation-Range Continuous Face-Sheet Micro Deformable Mirror Based on SOI
p.185

Towards Multi-Classification of Human Motions Using Micro IMU and SVM Training Process
p.189

Study of Microfluidic Chip with In Situ Prepared Chitosan Membrane Immobilized with Glucose Oxidase
p.194

Fabrication and Performance Analysis of PMMA Substrate Diffuser/Nozzle Micro-Pump by CNC Milling Machine
p.198

Effect of the Thickness Uniformity in Micro-Electroforming Process by Dividing Conductive Zone on Cathode
p.202

Single Neuron Adaptive PID Control of the Silicon-Based Integrated Micro Nano-Positioning XY-Stage
p.207

HomeAdvanced Materials ResearchAdvanced Materials Research Vols. 60-61Towards Multi-Classification of Human Motions...

Towards Multi-Classification of Human Motions Using Micro IMU and SVM Training Process

Abstract:

This paper introduces a novel approach for human motion recognition via motion feature vectors collected by A Micro Inertial Measurement Unit (µIMU). First, µIMU that is 56x23x15mm3 in size was built. The unit consists of three dimensional MEMS accelerometers, gyroscopes, a Bluetooth module and a Micro Controller Unit (MCU), which can transmit human motion information through a serial port to a computer. Second, a human motion database was setup by recording the motion data from the µIMU. The motions include fall, walk, stand, run and step upstairs. Third, Support Vector Machine (SVM) training process was used for human motion multi-classification. FFT was used for feature generation and optimal parameter searching process was done for the best SVM kernel function. Experimental results showed that for the given 5 different motions, the total correct recognition rate is 92%, of which the fall motion can be classified from others with 100% recognition rate.