Neural network for improve ORB-SLAM2 on XZ plane

The relevance of autonomous vehicles has increased in recent years, an important part of autonomous systems are SLAM systems which can reconstruct routes traveled and locate the vehicle within the created map. One of the best options for performing SLAM are vision systems. The accuracy achieved with this type of systems in motion calculation is below 2 percent error in translation, and different approaches have been used to reduce this error. In this work we propose a complement to the ORB-SLAM2 system to reduce the error in the translation calculation, a filter combined with a neural network is used to reduce the error. The error obtained with this plugin was 35 percent less than the original ORB-SLAM2 system.