OD-SLAM: Real-Time Localization and Mapping in Dynamic Environment through Multi-Sensor Fusion

Only when the robot knows its position can it execute some specific functions. But in many cases, the robot faces an unknown environment and doesn't know where it is. Visual SLAM system uses the camera on the robot to capture the image of the surrounding environment for positioning. This is a method that relies on external information for positioning. When the external information is dynamic, the pose estimated by the traditional visual SLAM system is inaccurate. So a novel SLAM system is proposed in this paper, which mainly uses a neural network and multi-sensor fusion to realize robot positioning in a dynamic environment. The SLAM system is developed based on ORB-SLAM2. It recognizes highly dynamic objects in the image through the neural network and removes them from the image. Because the neural network can't accurately identify dynamic objects every time, it will affect the pose estimation of the SLAM system. Therefore, the SLAM system proposed in this paper uses odometer to constrain features extracted in the image, so as to obtain static features, and finally estimate a more accurate pose. Finally, the experimental results show that our system has better real-time and accuracy compared with the traditional visual SLAM systems and other dynamic SLAM systems.