Multi-sensors Fusion Based 3D Object Tracking

3D object tracking is a fundamental problem in computer vision community,and has been research focus for lots of experts.Classic model-based tracking methods try to align the projected edges of a 3D model with the edges of the image.With image processing techniques,we got the contour of a 3D model projecting in the image.Then we got the point set by uniform sampling on the contour.Therefore,aligning edges is actually matching the sample points with their candidates along normal vector.However,wrong matches at low level cased by illumination changes,occlusions,fast motion etc.can make these methods fail.In order to solve wrong matches at low level caused tracking fail,in this paper we propose a new approach to improve 3D model-based tracking robustness.The main contributions are as follows:1.Camera rotation estimation via a combination of inertial measurement sensors.The units used in the inertial measurement sensors for rotation determination are a 3-axis gyroscope(detecting angular velocity),a 3-axis accelerometer(detecting the the direction of the earth's gravity field)and a 3-axis magnetometer to measure the direction of the earth magnetic field.We first use an Kalman filter on the raw data to reduce the measurement error.Angular velocity needs to be integrated once along with recurrence to obtain the camera's rotation.After that,we present a sensor fusion algorithm using complementary filter to improve the rotation accurate.Besides,accelerometer data is integrated once to obtain the camera's velocity,which is useful in procedure of camera translation estimation.2.Camera translation estimation via particle filter.We employ the obtained camera rotation and velocity to a particle filter,which estimate the camera pose using a dynamic model to establish a prediction and an observation model to correct it.Each particle represents a potential camera pose.As the camera rotation is settled,particles only have three degrees of freedom.Therefore,smaller state space allow us to use larger number of particles for guaranteeing tracking accuracy while limiting the computational cost.3.Complete an algorithm framework.In this paper,we integrate inertial measurement sensors and particle filter into one framework.In summary,we focus on 3D object tracking and propose a tracking method with sensor attachment.The tracker presented in this paper was tested on different video sequences and demonstrate its feasibility.Qualitative and quantitative analysis show that the proposed method performed successfully.