News

On 27 and 28 March, AI4CCAM took part in HyCHA’24 – Hybrid Artificial Intelligence Workshop: From knowledge and human integration to model explanation (Knowledge, Human and Machine Learning Hybridization), held in Gif-sur-Yvette (France), co-organised among others by SystemX, AI4CCAM partner.

AI4CCAM was represented by Arnaud Gotlieb, Simula, AI4CCAM coordinator who participated in the event with a poster.

Hybrid Artificial Intelligence (AI) broadly encompasses all approaches that combine several AI methods, whether symbolic or numerical, qualitative, semi-qualitative or quantitative. It is a field that is currently enjoying renewed interest, particularly because the hybridization is a way of addressing the respective weaknesses of different approaches and tackling certain current AI challenges such as trust and explainability.

In Automated Driving, scene understanding is a crucial issue. However, today’s explanations are based on quantitative analysis: Heatmaps, importance map, occlusion analysis, etc. But there is still no understanding of events chain and consequences. Given an action taken by an ego-car in a scene, the goal is to automatically explain the interactions between the actors that led to such action Decision Making Explanation Action.

AI4CCAM is based on Qualitative Reasoning.

Given a scene, we apply Qualitative Constraint Acquisition on each frame Then, we create the Qualitative eXplainable Graph (QXG) from sensor data.

Given a set of QXGs and actions (cruising, accelerating, etc.), we train one-class classifiers to give action-labels to each object relation chains. These action labels can be helpful to automatically document the driving scene.

Considering action-labels with relation chains, our method automatically proposes possible ego-car action explanations, by considering interactions between the car and each object.

User feedback on possible explanations allows us to strengthen the result and properly document automated driving scenes.

For further information on the event, click here

Check the project library for the Poster!