Monocular SLAM and Obstacle Removal for Indoor Navigation

Visual Simultaneous Localization and Mapping (SLAM) is one of the hot topics in computer vision. For the past few years, the AI and deep learning technology research have been widespread used in self-driving technology and surveillance system etc., gaining more and more attention from researchers and public media. The combination of AI technology and robot perception is inevitably going to be a trend. This paper aims at removing the obstacle to enhance the SLAM system performance that based on popular open source framework ORB-SLAM2 in dynamic environment. Moving objects will bring noise in camera pose estimation, besides, when in re-localization, the robot returns to the previous place finding the previous landmark mismatches because of its movement. The system will be confused and misdirected. A novel approach is proposed to remove the obstacle in real environment by using convolutional neural network (CNN) to generate a segmentation mask of obstacle object so as to eliminate the interference by moving object. Our experiment result shows an impressive outcome of practical use and benchmark dataset test.