Abstract:

Estimating camera pose in a scene with known 3D information is important in several fields such as autonomous driving, augmented reality and virtual reality. There are many methods to directly return to the camera pose from the input image, and then calculate the camera pose through the 3D coordinates of the returning pixels. However, the problem of these methods is that they are seriously coupled with the training scene and lack the generalization ability in the new environment. The convolutional neural network should focus on learning the robust and invariant visual features. Therefore, a direct alignment optimal method is proposed based on multi-scale features. It takes the feature similarity as the measurement form and the camera pose as the optimization quantity, and estimates the accurate 6 degree of freedom (6DOF) pose of the camera through end-to-end training from pixel to pose. The system separates the model parameters from the scene, has strong generalization ability for new scenes, and has good pose location accuracy.

Key words: multiscale visual feature, pose optimization, feature alignment, generalization ability