Development of a mapping system on an autonomous vehicle using a Fully Convolutional Neural Network and Fast SLAM Algorithm

There are many challenges when it comes to autonomous vehicle movement, one of which is developing an accurate and precise internal mapping system. Autonomous vehicles use internal maps to move from a starting point to destination point. Many methods are used in creating these maps, but because they still display weaknesses, further development is required. This research combines the FastSLAM 2.0 algorithm with a fully convolutional neural network (FCNN) model using the road features recognized by the FCNN algorithm as the object of observation of the FastSLAM 2.0 algorithm. This method was tested to form a map of the environment around the Faculty of Engineering, Sriwijaya University, Inderalaya Campus. In the training, the Adam optimizer and Adam combined with batch normalization (BN) model showed good accuracy: 82.07% and 78.08%, respectively. The application of this method succeeded in forming a map similar to Google Maps using the FCNN observation model. The map that was successfully formed had an IoU of 0.159 against the Google Maps map obtained with the Adam + BN model.