Visual odometry for on-road vehicles based on trifocal tensor

The accurate positioning is the core technology of mobile robot. The paper proposes a visual odometry method based on trifocal tensor to get the high-precision positioning information of autonomous robot. Two-wheel car was used to simulate the mobile robot, where monocular camera was mounted on. We employed camera calibration algorithm to get intrinsic parameters, the IPM (Inverse Perspective Mapping) to get the top view of pavement images, the improved SURF to detect and match feature points, trifocal tensor to calculate the fundamental matrix after outliner points removing based on RANSAC algorithm and calculate the car pose from the fundamental matrix. Finally, the Kalman filter was adopted to estimate the pose of the car. Experimental results and analysis demonstrate that visual odometry based on trifocal tensor well restrain the drift error of visual positioning method.