Error mitigation for untrained data in parking vehicle shape estimation by deep learning using millimeter-wave radar
- Delivery
- Available on this site
- Format
- Price
- Non-members (tax incl.):¥1,100 Members (tax incl.):¥880
- Publication code
- 20219086
- Paper/Info type
- Other International Conferences
- Pages
- 1-2(Total 2 p)
- Date of publication
- Sep 2021
- Publisher
- JSAE
- Language
- English
Detailed Information
Author(E) | 1) Tokihiko Akita, 2) Haruya Kyutoku, 3) Ukyo Tanikawa, 4) Yusuke Akamine |
---|---|
Affiliation(E) | 1) Toyota Technological Institute, 2) Toyota Technological Institute, 3) SOKEN, INC., 4) SOKEN, INC. |
Abstract(E) | Deep neural network (DNN) is capable of highly accurate recognition for trained data, but it cannot guarantee estimation error for untrained data. This is an essential generalizability challenge in machine learning, and it is difficult to solve fundamentally. To solve this problem, we embed the intrinsic knowledge of the model to be estimated into the training model to improve generalizability. The error model is trained for the estimation results. If the error is estimated to be large, we replace the estimation logic with a more stable deductive model. In this way, the maximum error can be suppressed. In this paper, we apply this method to the shape estimation of a parking vehicle using millimeter-wave radar. A parametric model with a minimized number of parameters is used to represent the shape of the vehicle, which is trained by a convolutional neural network (CNN) to estimate the shape parameters. If the error is judged to be too large, the positional error is corrected from the radar reflection pattern. As a result, it was confirmed that large errors could be suppressed for untrained data. |