Session: 05-04-01 CCUS and Underwater Development/ Utilization

Submission Number: 156199

Obstacle Detection and Avoidance Method Using Object Detection Model and Stereo Vision 

Autonomous underwater vehicles (AUVs) are vehicles that navigate underwater without cables and conduct observations. Since AUVs generally have limited means for operators to be involved during navigation, they are required to have higher levels of autonomy than vehicles that are remotely controlled using cables.

The basic function of AUVs is to detect and avoid obstacles. The former is broadly understood as the problem of the AUV's surrounding environment recognition, while the latter is understood as the problem of route planning for the mission.

Sensors such as sonar and optical cameras are used as a means for AUVs to recognize the surrounding environment while navigating. The measurement range and resolution depend on the sensor performance, but various research is being conducted on how to use the obtained sensor data.

We propose a method to recognize the surrounding environment by combining an object detection model and stereo vision. In this method, the object detection model is used to determine whether a detected object is known or unknown, and the stereo vision is used to determine the distance and size of the detected object. Based on the results of both, it is determined whether the object should be avoided.

There are various obstacle avoidance methods, but we propose a method to solve this problem as a route generation problem based on the planned route and detection information. Dijkstra's algorithm is a common route search algorithm, and the A* algorithm, which further improves search efficiency, is also available. These algorithms usually perform searches on a two-dimensional plane, but in this study, we use a method that extends the A* algorithm to three dimensions in order to apply it to AUVs navigating underwater.

The proposed method was evaluated in a simulation. The software under development is configured using ROS2, a platform for robot development, and is configured to facilitate mutual exchange between the simulation environment and the real environment. Therefore, a virtual seabed environment was prepared in the simulator software, and the proposed method was evaluated in a situation where several objects were placed.

As a result, it was confirmed that the system can change course to avoid only objects that are recognized as obstacles. It was also confirmed that the AUV's on-board computer can recognize the surrounding environment and generate routes in real time.

We plan to continue evaluating and improving the system in the future. We also plan to implement it on an actual AUV and evaluate it in actual sea trials.

Presenting Author: Kazuya Iwashita Japan Agency for Marine-Earth Science and Technology

Presenting Author Biography: Kazuya Iwashita
Engineer
Marine Robotics Engineering Group
Engineering Department
Japan Agency for Marine-Earth Science and Technology

B.Sc. in Mathematics, Faculty of Science, Chiba University

He is currently focused on the research and development of machine learning models for image recognition in underwater exploration.