Relative Pose Estimation of Visual SLAM Based on Convolutional Neural Networks

Pose estimation is the core module of the visual SLAM system, which is usually calculated by the front-end visual odometry. The calculated pose gives the initial estimation of the camera pose that is the basis for later map construction and path planning. The target of pose estimation is to calculate the relative motion between two frames. The absolute pose of the latter frame can be calculated according to the absolute pose of the previous frame and the relative transformation matrix of the two frames. The main methods for calculating the relative motion of two frames are feature-based method and direct method, and both of them have been successfully applied in various visual SLAM systems. But they all have the disadvantage of needing to carefully adjust the parameters, and sensitive to environment and illumination changes, image blur and other factors. With the great success of deep learning in the field of computer vision, researchers began to explore the use of deep learning methods to solve pose estimation problems. This paper proposes a neural network model for calculating the relative motion of two frames. The training and testing are carried out on a public dataset. The experimental results show that the model shows good performance for known scenes and has certain generalization ability for unknown environments, proving that deep learning methods have the potential to solve the problem of pose estimation.