Development of VPS for Indoor UAV System: Vision based SLAM and Pedestrian Tracking

Abstract

This paper aims to develop a vision-based positioning system for indoor UAVs. Indoor Unmanned Aerial Vehicles(UAVs) are usually very small, and their load limit is harsh. Also, GPS cannot be used to operate indoors. So, it is essential to position them using vision or radar because it is difficult to locate them using sensors such as GPS. Vision-based positioning is more advantageous than lidar-based positioning due to power and weight issues. With edge computing, the lightest and lowest power system can be configured using monocular camera positioning. To this end, we propose a monocular camera-based depth estimation method, develop a vision-based positioning system for indoor UAVs using this method, and experiment with Simultaneous Localization And Mapping(SLAM) and pedestrian following using monocular depth estimation of the UAV system. To develop monocular depth estimation, we modified the decoder of the neural network-based monocular depth SOTA method GLPdepth [1] using atrous spiral pyramid(ASPP) and Simplegate [2], and the experimental results showed a 1.1% performance improvement over the GLP-depth estimation method on the NYU depth V2 dataset. Based on this neural network, we compared the performance of monocular vision SLAM and stereo SLAM. As a result, we confirmed that it showed about 85% performance compared to binocular cameras in the XY axis. In addition, in the case of a UAV following a person, we confirmed that we could control the indoor UAV to fly at a certain distance from the person by fusing monocular depth estimation, Yolo-V3 [3] based person recognition, and stereo camera-based RTAB-Map SLAM to estimate the location of the person.