Abstract:

Visual localization technique is an important component in the field of visual navigation and autonomous driving. A camera 6 degree of freedom(6DOF) localization method based on geometrically constrained Siamese convolutional networks is proposed, which uses a geometrically constrained way for the convolutional network to learn the relative positional relationship between the query image and the reference image to obtain the global position of the query image. Meanwhile, this method uses a high-performing backbone feature extraction network and a multi-task joint loss function training strategy simultaneously to further improve the localization accuracy, stability, and generalization of the proposed method. The feature distance metric loss function is also designed to enhance the differentiation of similar images. The validation data on indoor and outdoor public datasets show that the proposed method is more competitive when it is compared with similar methods.

Key words: camera location, Siamese convolutional network, relative pose, generalization capability, pose accuracy