Kheriji, Walid (Vedecom), Hadj Selem, Fouad (VEDECOM), Ben nejma, Ghada (Vedecom), Bontemps, Thomas (Vedecom), Durville, Laurent (Vedecom), Rahal, Mohamed-Cherif (VEDECOM)

Extracting Scenarios for Automated Driving Systems: A Statistical Based Occupation-Grid Approach with Deep Semi-Supervised Learning

Scheduled for presentation during the Regular Session "Generating driving scenarios I" (ThBT11), Thursday, September 26, 2024, 15:10−15:30, Salon 19/20

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on October 7, 2024

Abstract

Certification of advanced driver assistance systems (ADAS) is mainly based on scenario testing. Finding all possible road scenarios and learning their statistical distribution is fundamental and always based on examples extracted from real driving data, which is a huge challenge. Existing methods for extracting driving scenarios include both rule-based and machine-learning approaches. While rule-based approaches suffer from scalability issues, machine learning approaches encounter limitations due to the format of scenario data, characterized by variable-length time series with dynamic variations in vehicle perception. More specifically, recurrent neural networks and transformer approaches are confronted with unstable data formats and quality because of perception problems and variations in the number of obstacles. Convolutional neural networks approaches introduce occupancy grids to solve data format issues, but lack statistical justification for transforming time series into images.

In this paper, we present a new approach to transform driving data into tensors, based on a rigorous statistical process that ensures stability and scalability compared to existing methods limited to specific formats or scenarios. Our approach aims to standardize the presentation of scenarios in a uniform way to enable the use of recent AI approaches for large-scale scenario collection. This will help to detect both common and rare scenario classes. In addition, we design an efficient learning strategy adapted to the volume of data available, their state of annotations, and the intrinsic characteristics of deep learning models. Furthermore, we validate our methodologies on real European driving data.