|
A Mission Allocation Algorithm Using Multi-Agent Reinforcement Learning |
|
Dasol Lee(Agency for Defense Development, Korea) |
| This paper describes initial research results on a mission allocation algorithm using multi-agent reinforcement learning (MARL). Each agent selects a task through a learned network, and sequential execution of tasks leads to the given mission completed. StarCraft II real-time strategy game is utilized as an environment for MARL, and learned results show that MARL framework can be applied for mission allocation effectively. |
|
|
Depth Image Thresholding for Tomato Detection in the Greenhouse |
|
Dasom Seo, Kyoung Chuhl Kim, Kyungdo Kwon, Yeji Kang, Gyeongho Moon(National of Agricultural Sciences, Rural Development Administration, Korea) |
| The importance of agricultural intelligence and automation has been increased recently. In particular, study on fruit-harvesting robots has been actively conducted since it can save labor in the green house. This paper presents a three-dimensional image analysis for tomato detection in an automatic harvesting robot. Region segmentation is required because tomatoes in other lines as well as in the target line can be captured in the same frame. Depth information is acquired by capturing tomatoes with an RGB-D camera. The region further than the distance between the camera and the target line should be removed from the frame image. |
|
|
Real-time Simulation Model for an Under-water Construction Robot using the Recursive Subsystem Synthesis Method |
|
Chang-Ho Lee, Sung-Soo Kim(Chungnam National University, Korea) |
| In this paper, a real-time simulation model is presented for an underwater construction robot (UWCR) simulator using the recursive subsystem synthesis method (RSSM). To apply RSSM, the UWCR is decomposed into several open-loop and closed-loop subsystems. Equations of motion can be generated for each subsystem. For the entire UWCR simulation, the dynamics effects of a subsystem are recursively transferred to the other subsystem. The UWCR real-time model has been realized and validated with the multibody software RecurDyn. The efficiency of the model has been also investigated according to the different integration step-size. |
|
|
Development of Steerable Needle Segmentation from Ultrasound Images using Deep Learning Algorithm |
|
Bijoya Lala, Seong Young Ko(Chonnam National University, Korea) |
| For the needle steering system, real‐time needle shape or tip position is an important information. In this study, we proposed a method based on the deep neural network to segment and track needle’s shape in two-dimensional (2D) ultrasound (US) images. We collected 2D US images using a commercial US system after inserting an 18-gauge needle into soft tissue. A frame grabber was used to capture the image from US system and to transmit it to the computer, and eventually the needle was segmented and visualized. Experiments on different tissue types and different needle insertion paths showed its effectiveness and feasibility to be used for the needle tip tracking. |
|
|
Channel Fusion Method for Object Detection of Multi-Spectral Image |
|
Youngjin Kim, Jonghee Park, Minyong Sung, Sung-Joon Jang(Korea Electronics Technology Institute, Korea) |
| In this study, we used multi-spectral image (RGB, IR) as an input of object detection for deep learning. Our multi-spectral image was captured at the same time and the same point of view, but there is difference in visibility. Therefore, for multi-spectral images, channel-wise calibration should be preceded to use as an input of deep learning while strengthen its channel characteristic. |
|
|
A Spectral Image Analysis on the Detecting Wasp Nest using by Drone |
|
Kyoung Chuhl Kim, Dasom Seo, Inchan Choi, Kyungdo Kwon, Young Ki Hong(National of Agricultural Sciences, Rural Development Administration, Korea) |
| The Beekeeping Farming industry is struggling as the number of honeybee is decreasing due to the increase in Wasps. To solve the problem, It is necessary to develop a system to detecting the wasp nest. Drone is great for exploring large areas. For Wasp nest detection. We installed a spectral sensor on the drone. The main images acquired by drones are leaves, branches and wasp nest. We analyzed the spectral characteristics for classification. As a result of the analysis, characteristics of 780, 840nm for leaves, 580nm for dry branches and 620nm for wasp nest were derived. Through this study, We are developing to detecting system of wasp nest using by drone. |
|