A Semi-dense Visual Odometery Based on 3D Scene Flow

Abstract

Aiming at the problem of how to accurately locate the intelligent industrial robots in an indoor dynamic environment, a direct method of semi-dense visual odometery algorithm based on 3D scene flow is proposed. It divides the scene into static parts and dynamic parts. The static part is used to solve the camera motion, and the dynamic part is used to estimate the scene flow of the moving object. By using the inverse depth instead of the depth of the geometric error to parameterize the geometric error and optimize both the photometric error and the geometric error at the same time, it provides a better fit for the geometric error distribution and the optimization function. In the comparison of some mainstream visual odometry methods, the experimental results under different standard data sets show that, under indoor dynamic environment, the algorithm can more accurately estimate the pose of intelligent industrial robots in real time and improve the robustness of the system.