Research On Visual Attention Tracking Method Based On Multi-scale Inertial Sensor Fusion

Visual attention tracking is widely used in virtual/augmented reality,rehabilitation medicine,self-service and other scenarios.The traditional visual mode may generates many constraints in movable scenes,such as collision after wearing glasses.Using a headphone with a built-in inertial sensing unit,the problem of visual attention tracking can be transformed into perception of head movement direction.The lack of a magnetometer in the headphone due to interference from speaker magnets adds a significant challenge to accurately tracking head movements,i.e.the initial orientation of the head cannot be obtained in the absence of a magnetometer,and the error of the head in the horizontal plane cannot be corrected;In addition,the user wears the headphone in different orientations/angles,which causes the headphone coordinate system to be inconsistent with the head coordinate system.In view of the above deficiencies and requirements,the paper focuses on the accurate visual attention tracking method.The main work can be summarized as follows:First,a multi-scale inertial sensing coordinate system transformation method is proposed.Through the projection modeling of the headphone and the head coordinate system,the rotation matrix from the headphone coordinate system to the head coordinate system is obtained;Through the projection modeling of the mobile phone and the head coordinate system,the initial direction the user is facing is obtained.Second,design and implement a multi-scale inertial fusion visual-attention tracking system.Based on the observation that "the range of human head rotation is very small",the system fuses the rotation model of the head into the particle filter to make up for the defect that the inertial sensor in the headphone cannot correct the horizontal direction error.Finally,experiments are carried out in the real environment,and the experimental results are analyzed.The results show that the median error of visual attention tracking is about 5.6° or 8.2° when the user is silent or generates micro-scale action disturbance.