Realization of robust SLAM in non-artificial environment based on feature extraction function of deep learning

2019 Fiscal Year Final Research Report

• Project/Area Number: 17K06485
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Control engineering/System engineering
• Research Institution: Hokkaido University
• Principal Investigator: EMARU TAKANORI 北海道大学, 工学研究院, 准教授 (30440952)
• Project Period (FY): 2017-04-01 – 2020-03-31
• Keywords: SLAM

Outline of Final Research Achievements

The purpose of this research project is to realize robust SLAM using feature points which are automatically extracted by deep learning in order to perform highly accurate self-localization in a non-artificial environment represented by agriculture and forestry fields. The main research results are as follows: 1) To identify weeds and crops, we proposed a multi-modal deep learning system that utilizes temperature information in addition to RGB and Depth. 2) We proposed image processing algorithm to efficiently acquire a large amount of data used for deep learning. 3) In order to realize a system that can autonomously drive at low cost in the agriculture field, we constructed a system that recognizes the environment using an inexpensive Depth sensor and verified the effectiveness by simulation and actual machine experiments.

Free Research Field

ロボット工学

Academic Significance and Societal Importance of the Research Achievements

これまでSLAMに関する研究は屋内・屋外人工環境で行われており，農林業環境に代表される非人工環境下でロバストに実現することは未だ行われていないため学術的・社会的にも非常に意義のある研究である。本研究を遂行する上で重要な要素技術となっている作物・樹木の3次元座標獲得および特徴点抽出は，ロバストなSLAMを実現すると同時に除草・収穫作業を効率的に行うための重要な情報となり，高機能農林業機械開発の中核的な技術となり得る。また個々の作物・樹木の情報を含むロバストな地図情報は，センサネットワークやビックデータ解析と組み合わせることにより飛躍的に情報の質を向上させ，同時に生産性の大幅な向上が期待できる。