2023 Fiscal Year Final Research Report

New Approach to Deep Learning by Introducing Language Model for Drastic Improvement of Automated Driving Reliability

Research Project

PDF

Project/Area Number	21K03985
Research Category	Grant-in-Aid for Scientific Research (C)
Allocation Type	Multi-year Fund
Section	一般
Review Section	Basic Section 20020:Robotics and intelligent system-related
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)	2021-04-01 – 2024-03-31
Keywords	深層学習 / 半教師あり学習 / Variational Autoencoder / 能動学習 / ミリ波レーダ / 駐車車両形状推定
Outline of Final Research Achievements	A semi-supervised learning method was created to improve the accuracy of parking vehicle shape reconstruction for unlearned data. It is difficult to generate a large number of correct shapes of parking vehicles for measured real-world data, resulting in a loss of accuracy. In contrast, this method improves accuracy with only a small amount of supervised training data. This was achieved by combining pseudo correct value generation, active learning, and consistency regularization methods. For the generation of pseudo correct values, we created a method that uses self-supervised learning with VAE and a method that determines the consistency of the estimated shape with the model of the correct shape. With this method, we confirmed that with only 16% of the supervised data, the accuracy of estimating the shape of a parking vehicle is equal to or better than the marginal performance at 80%.
Free Research Field	情報科学
Academic Significance and Societal Importance of the Research Achievements	深層学習の認識精度は学習データに依存する。そのために大量の正解値が必要であるが、コストの課題がある。本研究では少ない正解値で、大量の正解値付与の場合と同等の性能を得られることを示し、課題解決策を提案した。特に物体形状推定における半教師あり学習の手法は研究例が少なく、新しい可能性を示すものである。ここではミリ波レーダを用いた駐車車両形状復元を対象とした。ミリ波レーダは耐環境性が最も高いが、分解能が低くノイズが多いため、画像より遥かに正解値生成コストがかかるため、これを低減することは特に有効である。ソナーによる駐車支援をミリ波レーダ流用に置き換えることにより、低コスト化と意匠改善できる。