Multimodality Weight and Score Fusion for SLAM

Abstract

Simultaneous Localization And Mapping (SLAM) is used to predict the trajectory by the Autonomous Navigation Robots (ANR), for instance Self-Driving Cars (SDC). It computes the trajectory through sensing the surroundings, like a visual perception of the environment. This work focuses on the performance improvements of a SLAM model using multimodal learning: (i), early fusion via layer weight enhancement of feature extractors, and (ii), late fusion via score refinement of the trajectory (pose) regressor. The comparative analysis on Apolloscape dataset shows that the proposed fusion strategies improve localization performance significantly. This work also evaluates applicability of various Deep Convolutional Neural Networks (DCNNs) for SLAM.