Disparity Estimation Based on Feature Map Correlation using Contrastive Learning

Takeru Ninomiya･Takeshi Endo･Hideaki Kido (Hitachi)･Kota Irie (Hitachi Astemo)

Depth information can be estimated with high accuracy using deep learning. However, for camera pairs with long baseline, the accuracy of disparity is reduced because the visual difference between left and right images. In this paper, we proposed disparity estimation method based on feature map correlation using contrastive learning. By taking into account visual differences between the left and right images, we improve the accuracy of disparity estimation. Experimental results show that the proposed method improves accuracy within a neighborhood of 20 m.

Research on Small Object Detection on Deep Learning in Autonomous Driving

Enhi Sen (dSPACE Japan)･Ryota Yagami (Nagaoka University of Technology)

In the verification of autonomous driving and its algorithms, it is very important to detect small objects in the distance, but it is difficult to accurately detect the small objects themselves and the specified size. Therefore, this research uses SAHI, which can make inferences without scaling the images down, GAN, which can enlarge images, and fine-tuning with small object dataset, to accurately detect objects up to a specified size.

Study on Method of Development of Driving-Decisions Model for Automated Bus Traveling Fixed Route

Daichi Minagawa･Manabu Omae (Keio University)

In this presentation, we propose a method to efficiently construct an inference model for driving decisions by training a model that outputs the required driving decision results according to the location of an automated shuttlebus, based on images collected during route travel. We confirmed that this method is effective in adding and expanding the function of driving decisions when automating buses that travel fixed routes.

Necessity of an Estimation Function of Driver's Driving Willingness in Autonomous Driving Systems

Toshiya Arakawa･Kazuya Itakura (Nippon Institute of Technology)･Jun Tajima (Misaki Design)

We discusses the necessity of a function for understanding whether or not the driver has the willingness to drive a vehicle in a Level 2 or higher autonomous driving system. We also construct a discriminator of the driver's willingness to drive by machine learning, and evaluate the validity of the estimation of the driver's willingness to drive by the constructed discriminator.