Depth Estimation and Object Detection for Monocular Semantic SLAM Using Deep Convolutional Network

Abstract

It is still challenging to efficiently construct semantic map with a monocular camera. In this paper, deep learning is introduced to combined with SLAM to realize semantic map production. We replace depth estimation module of SLAM with FCN which effectively solves the contradiction of triangulation. The Fc layers of FCN are modified to convolutional layers. Redundant calculation of Fc layers is avoided after optimization, and images can be input in any size. Besides, Faster RCNN, namely, a two-stage object detection network is utilized to obtain semantic information. We fine-tune RPN and Fc layers by transfer learning. The two algorithms are evaluated on official dataset. Results show that the average relative error of depth estimation is reduced by 12.6%, the accuracy of object detection is improved by 10.9%. The feasibility of the combination of deep learning and SLAM is verified.