數據驅動的高擬真度自動駕駛系統測試情境生成

Abstract

本研究針對自動駕駛系統(Automated Driving System, ADS)的測試，提出了一套從真實世界資料中提取有用資訊，轉化為測試情境，並建立模擬平臺的方法。流程包括數據預處理、車輛行為篩選、車輛行為資料分析和情境模擬呈現等步驟。這種方法不僅能彌補文獻中使用速度定值產生測試情境的限制，而且具有通用性，可適用於不同地區。考慮到車輛行為的複雜程度，本研究以車輛跟隨和車道變換作為示範，分別使用簡單篩選和深度學習的方式提取資料集中特定的車輛行為。在資料分析階段，能獲取不同地區的參數範圍，有助於生成具覆蓋性或專注於單一速度區間的測試情境。這些分析結果有助於更全面地瞭解並針對不同地區進行更準確的情境生成。情境模擬則結合OpenSCENARIO 格式，實現在不同測試軟體間的兼容性並提高情境測試的效率。總的來說，本研究結合數據驅動，提供了一個完整的情境生成流程，有望朝著高擬真度測試情境的方向發展，透過測試情境找出ADS 的缺陷和潛在風險。

This study proposes a methodology for testing Automated Driving Systems (ADS) by extracting useful information from real-world data, converting it into test scenarios, and applying it to a simulation platform. The process includes data preprocessing, vehicle maneuver filtering, vehicle maneuver data analysis, and scenario simulation presentation.This approach not only addresses the limitations of regulatory testing scenarios generated with constant speeds but also offers general applicability to different regions. Considering the complexity of vehicle maneuvers, this study demonstrates vehicle following and lane changing maneuvers by using simple filtering and deep learning methods to extract specific vehicle maneuvers from the dataset. During the data analysis phase, obtaining parameter ranges from different regions helps in generating test scenarios that are either comprehensive or focused on a single speed range. These analytical results contribute to a more thorough understanding and accurate scenario generation for different regions.The scenario simulation integrates the OpenSCENARIO format, ensuring compatibility across different testing software and enhancing the efficiency of scenario testing. Overall, this study combines data-driven approaches to provide a complete scenario generation process, aiming towards high fidelity testing scenario.