Improved CNN-based Path Planning So an Autonomous UAV Can Climb Stairs By using a LiDAR Sensor

Abstract

Unmanned aerial vehicles (UAVs) have tremendous potential in civil and public areas. These are especially beneficial in applications where human lives are threatened. Autonomous navigation in unknown environments is a challenging issue for UAVs where decision-based navigation is required. In this paper, a deep learning (DL) approach is presented that aids autonomous navigation for UAVs in completely unknown, GPS-denied indoor environments. The UAV is equipped with a monocular camera and a light detection and ranging (LiDAR) sensor to determine each next maneuver and distance calculation, respectively. For deeper feature extraction, a version of You Only Look Once (YOLOv3-tiny) is improved by adding a convolution layer with different filter sizes. The process is observed as an exercise where the DL model classifies the targeted image as stairs or not stairs. We created our dataset considering the indoor scenario for specific implementation. Comprehensive experimental results are compared with YOLOv3-tiny, indicating better performance in terms of accuracy, recall, F1-score, precision, and maneuvering movements.