깊이 이미지를 이용한 딥러닝 기반 충돌 확률 예측 및 장애물 회피 알고리즘

Abstract

This paper proposes a vision-based navigation algorithm in cluttered environments by using a convolutional neural network (CNN)-based deep learning and depth images. Existing approaches for robot navigation are developed based on obstacle detection or an accurate map to generate a path, which requires complex computation. To solve this issue, we propose a deep-learning-based navigation algorithm that does not require a preobtained map or a recognition of obstacles. In addition, by exploiting a collision probability prediction using a depth image, our method automatically adjusts the velocity command while passing the complex area, so we can improve the safety of a mobile robot. To validate the performance of our proposed method, we conduct a simulation with a mobile robot in a Gazebo environment. The simulation results show that our method outperforms the conventional avoidance algorithm.