2021 年 39 巻 5 号 p. 471-474
For the purpose of Self-Localization by 3D-maps and camera images, we have worked on the realignment task that has bigger error than conventional DNN that matches between 3D-maps and camera images. To resolve this task, we payed attention the registration characteristics of CalibNet that realign them focusing on the axis of height of the vehicle. We devised the error given to training data for CalibNet, and we can train the network that can registration focusing on the side and depth axes of the vehicle. Combining with CalibNet and our network and doing two steps correction, we achieved that improving error by 1.4 times in average of each axes of the vehicle when the center error of 3D maps and camera images is max 55[cm].