Semi-Bionic SLAM Based on Visual Odometry and Deep Learning Network

Simultaneous localization and mapping (SLAM) is of great significance for robot navigation. Many SLAM systems have been developed based on biological cognitive mechanisms. In this paper, we propose Semi-Bionic SLAM (SB-SLAM), a framework which combines visual odometry based on the feature points with biological cognitive mechanism model. First, we build the head direction cell and place cell model through continuous attractor network to enhance computational efficiency and accuracy. Then we perform loop closure detection through a deep learning network. At the same time, a map update method is utilized to update the experience map. This mapping method can quickly and effectively detect the loop closure points in the map and correct the odometry drift in time. Compared with the existing biological mapping methods, the key cells model is retained instead of the complex biological system model to obtain lower elapsed time and practicability. Compared with ORB-SLAM2, the accuracy of the map is improved, and the speed of loop closure detection is accelerated.