학술논문
동적 장애물 회피를 위한 비전 기반의 Local Costmap 생성에 관한 연구 / Study for vision based local costmap for dynamic obstacle avoidance
Document Type
Dissertation/ Thesis
Author
Source
Subject
Language
Korean
Abstract
제 목 : 동적 장애물 회피를 위한 비전 기반의 Local Costmap 생성에 관한 연구 인공지능의 발달 및 비대면 서비스의 증가로 물류 산업에서부터 서비스업까지 다양한 분야에서 지능형 이동로봇의 사용이 증가하고있다. 이에 따라 사람과 로봇의 협업이 증 가하며 동적 장애물이 산재한 환경에서 로봇이 운용되는 경우가 많다. 오늘날 로봇은 Lidar 센서를 이용하여 주변 환경을 인식하며 물체의 움직임을 고려하지 않기 때문에 쉽게 고립되거나 심한 경우 사람과 충돌하기도 한다. 이러한 상황 속에서 로봇의 동적 장애물 회피 주행은 여러 산업들의 효율성 증대와 직결되는 중요한 문제이다. 본 연구 에서는 이러한 문제를 해결하기 위하여 비전 기반의 알고리즘들을 이용하여 물체의 움 직임이 고려된 Local Costmap을 생성하여 동적 장애물 회피 성능을 올리고자 한다. 카메라로부터 주변 환경을 인식하기 위하여 Semantic Segmentation을 통해 이미지를 분할하였다. 이 때 Semantic Segmentation을 완전히 추종하기 보다는 각 픽셀 당 장애물 존재 확률을 구했다. 또한 동적 장애물과 정적 장애물, 주행가능영역을 구분하여 주변 환경에서 동적 장애물의 위치를 파악하였다. 이미지로부터 정확한 2차원 좌표를 얻기 위하여 Inverse Perspective Mapping을 이용한 Bird's Eye View를 이용하였다. Semantic Segmentation으로 얻은 주행 가능영역은 z축 성분이 0이므로 Bird's Eye View로 변환 시 정확한 x,y 좌표에 픽셀을 매칭할 수 있다. 이 두 알고리즘을 이용하여 2차원 위치에 대 한 장애물에 대한 확률 분포를 얻을 수 있다. 동적 장애물의 움직임을 파악하기 위해서는 Object Tracking 알고리즘을 이용하였다. 각 이미지에서 동적 장애물을 추적하여 그 움직임을 계산한 후 이를 Bird's Eye View에 표시하여 동적 장애물의 2차원 속도를 추정하였다. 이 2차원 속도를 이용하여 위에서 구한 동적 장애물 확률을 이동 방향으로 증폭시켰다. 마지막으로 더욱 안정적인 주행을 위하여 장애물 충돌 확률을 팽창시킨 후 Local Costmap으로 사용하였다. 만들어진 Local Costmap을 이용한 주행을 위하여 DWA 알고리즘과 연결하여 지역경로 를 얻어 주행하였다. 동일한 환경에서 Lidar 기반의 Local Costmap을 이용했을 때 보다 더욱 유리한 회피 경로를 선택하는 것을 확인 하였으며, 동적 장애물 환경에서 충돌없 이 빠르게 목적지까지 주행함을 확인하였다.
Study for vision based local costmap for dynamic obstacle avoidanceWith the development of artificial intelligence and the increase in non-face-to-face services, the use of intelligent mobile robots is growing in various fields, from the logistics industry to the service industry. Consequently, collaboration between humans and robots is increasing, and robots are often operated in environments with scattered dynamic obstacles. Today's robots use Lidar sensors to perceive the surrounding environment, but they do not consider the movement of objects. Therefore, they can easily become isolated and, in severe cases, collide with people. In such situations, dynamic obstacle avoidance navigation for robots is a critical issue directly related to the efficiency enhancement of various industries.In this study, we aim to address these challenges by utilizing vision-based algorithms to create a local costmap that considers the movement of objects, thereby improving dynamic obstacle avoidance performance.To perceive the surrounding environment from the camera, we segmented the images through semantic segmentation algorithm. At this point, instead of completely adhering to semantic segmentation, we calculated the probability of the presence of obstacles for each pixel. Additionally, we differentiated between dynamic obstacles, static obstacles, and traversable region to identify the positions of dynamic obstacles in the surrounding environment. To obtain accurate 2D coordinates from the image, we used Inverse Perspective Mapping to create a Bird's Eye View. The traversable region obtained through semantic segmentation has a z-axis component of 0, so when transformed into Bird's Eye View, we can accurately match pixels to the exact x, y coordinates. By using these two algorithms, we can obtain a probability distribution for obstacles in 2D positions.To track the movement of dynamic obstacles, we used object tracking algorithms. We tracked dynamic obstacles in each image, calculated their movement, and then displayed it on the Bird's Eye View to estimate the 2D speed of dynamic obstacles. Using this 2D speed, we amplified the probability of dynamic obstacles obtained from above in the direction of movement. Finally, to achieve more stable navigation, we inflated the probability of obstacle collisions and used it as a local costmap.For navigation using the created local costmap, we connected it with the DWA algorithm to obtain a local path for navigation. We confirmed that choosing an avoidance path was more advantageous than when using Lidar-based local costmap in the same environment. Furthermore, we verified that the robot could navigate to the destination quickly without collisions in a dynamic obstacle environment.
Study for vision based local costmap for dynamic obstacle avoidanceWith the development of artificial intelligence and the increase in non-face-to-face services, the use of intelligent mobile robots is growing in various fields, from the logistics industry to the service industry. Consequently, collaboration between humans and robots is increasing, and robots are often operated in environments with scattered dynamic obstacles. Today's robots use Lidar sensors to perceive the surrounding environment, but they do not consider the movement of objects. Therefore, they can easily become isolated and, in severe cases, collide with people. In such situations, dynamic obstacle avoidance navigation for robots is a critical issue directly related to the efficiency enhancement of various industries.In this study, we aim to address these challenges by utilizing vision-based algorithms to create a local costmap that considers the movement of objects, thereby improving dynamic obstacle avoidance performance.To perceive the surrounding environment from the camera, we segmented the images through semantic segmentation algorithm. At this point, instead of completely adhering to semantic segmentation, we calculated the probability of the presence of obstacles for each pixel. Additionally, we differentiated between dynamic obstacles, static obstacles, and traversable region to identify the positions of dynamic obstacles in the surrounding environment. To obtain accurate 2D coordinates from the image, we used Inverse Perspective Mapping to create a Bird's Eye View. The traversable region obtained through semantic segmentation has a z-axis component of 0, so when transformed into Bird's Eye View, we can accurately match pixels to the exact x, y coordinates. By using these two algorithms, we can obtain a probability distribution for obstacles in 2D positions.To track the movement of dynamic obstacles, we used object tracking algorithms. We tracked dynamic obstacles in each image, calculated their movement, and then displayed it on the Bird's Eye View to estimate the 2D speed of dynamic obstacles. Using this 2D speed, we amplified the probability of dynamic obstacles obtained from above in the direction of movement. Finally, to achieve more stable navigation, we inflated the probability of obstacle collisions and used it as a local costmap.For navigation using the created local costmap, we connected it with the DWA algorithm to obtain a local path for navigation. We confirmed that choosing an avoidance path was more advantageous than when using Lidar-based local costmap in the same environment. Furthermore, we verified that the robot could navigate to the destination quickly without collisions in a dynamic obstacle environment.