Systematic Approach to Define Operational Design Domain from Individual Traffic Scene for Automated Driving Systems

Keisuke Shimono･Mitsuaki Hagino･Kimihiko Nakano (The University of Tokyo)

Explainable ODD to local government and citizens is required in the implementation phase of automated driving system (ADS). Each traffic scene is considered to design and development for ADS should be summarized systematically. Traffic scene includes several elements, therefore, those elements can be analyzed by graph approach to explain coverages of defined ODD to each traffic scene along the planned service route.

A Study on How Training Data Quality Affects the Performance of a VQA-based Model for Driving Scene Retrieval

Sota Nakanishi･Kento Ohtani･Kazuya Takeda (Graduate School of Informatics, Nagoya University)

In the field of autonomous driving, it has been pointed out that the quality of training data can adversely affect a model’s generalization performance and output stability.
This study focuses on a learning method that uses Visual Question Answering . We test this method using datasets that contain label bias and label inconsistency, and examine how stable the learning is and whether it avoids learning incorrect patterns.

Validation of autonomous vehicle management using remote assistance

Yasuhiro Akagi･Ryo Kanamori･Takayuki Morikawa (Nagoya University)

The remote assistance is a method which a human operator gives advice to an autonomous driving system to change its driving behavior from a remote location. In this paper, we impliment a remote assistance system to give advice for each traffic scenarios, such as passing through an unsignalized intersection, crossing a pedestrian crossing, and avoiding parked vehicles. We report on the effectiveness of the system based on field experiments in actual urban areas.

Occluded Vehicle Presence Estimation via Observable Vehicle Behaviors for Intersection Motion Planning

Koki Morita (Graduate School of Infomatics, Nagoya University)･Eijiro Takeuchi (TIER IV, inc.)･Kazuya Takeda (Graduate School of Infomatics, Nagoya University)

In autonomous driving, safe and efficient decision-making under occluded conditions is a critical important task.
overly conservative behavior that assumes the presence of unseen vehicles can cause unnecessary delays in traffic flow. We propose a method that observes the actions of visible surrounding vehicles and applies constraints to phantom vehicles in occluded regions, enabling more efficient motion planning. Using NuScenes dataset, real-world driving data, we demonstrate that our risk-constraint approach achieves more efficient decision-making than conventional models based on worst-case assumptions.