2021 Fiscal Year Final Research Report

Research of precise environmental perception method using millimeter-wave radar for all-weather autonomous driving

Research Project

PDF

Project/Area Number	19K21072
Project/Area Number (Other)	18H05897 (2018)
Research Category	Grant-in-Aid for Research Activity Start-up
Allocation Type	Multi-year Fund (2019) Single-year Grants (2018)
Review Section	0301:Mechanics of materials, production engineering, design engineering, fluid engineering, thermal engineering, mechanical dynamics, robotics, aerospace engineering, marine and maritime engineering, and related fields
Research Institution	Toyota Technological Institute
Principal Investigator	Akita Tokihiko 豊田工業大学, 工学(系)研究科(研究院), 特任上級研究員 (20564579)
Project Period (FY)	2018-08-24 – 2022-03-31
Keywords	ミリ波レーダ / 深層学習 / LSTM / CNN / VAE / クラス識別 / 形状推定 / 駐車
Outline of Final Research Achievements	A robust recognition method of the driving environment using millimeter-wave radar was created, and its effectiveness was shown using real-world data. Using bidirectional LSTM and time-series reflection maps, we constructed a method to classify cars, bicycles, and pedestrians from radar reflection signals only, and showed that it can classify them with higher accuracy than various other methods. We constructed an original deep learning network to estimate the type and shape of obstacles such as cars, fences, and curbs in various parking scenes in urban areas, and showed that it can estimate them with high accuracy using data measured in real environments. To improve generalizability to untrained data, we created a method using a model with vehicle shape knowledge and VAE to discriminate trained data, and showed that the maximum outline error can be suppressed.
Free Research Field	情報科学：自動運転の環境認識
Academic Significance and Societal Importance of the Research Achievements	ミリ波レーダは、照明や天候など悪環境下で頑健に走行環境認識が可能である。しかし、分解能が低くノイズが多いと言う課題があり、物体の種別識別や形状推定が困難である。独自の深層学習構成を創出してこの課題を改善し、高精度にこれらの機能が実現できることを定量的に示した。深層学習は、学習データに類似した入力に対しては高精度に認識可能であるが、未学習の入力では認識結果が予想外に逸脱することが課題である。これに対して、未学習データを判定して修正する手法を創出し、大きな誤差を抑制できることを定量的に示した。上記研究成果は、どの様な環境でも高い安全性が要求される自動運転や高度運転支援システムの実現に寄与できる。