Complex And Diverse Situation Modeling And Analysis Of Traffic Scenes For Developing Autonomous Driving Datasets

With the recent development of data-driven machine learning technologies,autonomous driving(AD)has attracted huge amounts of research and industrial attention.Most public are concerned about the reliability and safety of AD applications.For realworld applications,AD systems should be designed to deal with complicated scenarios and extreme weather situations to guarantee the reliability and safety.Therefore,how to quantitatively evaluate those challenging traffic scenes and how to construct training AD datasets by selecting most informative data samples have become central issues for developing AD perception and decision technologies.The goal of this thesis is to develop a set of AD situation evaluation metrics,computing methods,and evaluation benchmarks.The thesis mainly aims to resolve the following technological challenges.How to define intrinsic metrics to describe the AD situations in terms of complexity and diversity? How to construct a scene model to represent the multi-modal sensory information for AD systems? How to use these metrics and models to evaluate AD scenes and multimodal data samples?In this work,we propose a set of situation complexity and diversity metrics,multimodal scene models,graph-based scene models,and scene metric computing methods.Situation complexity and diversity metrics are used to describe the degrees of complexity and diversity of scenes.Multi-modal scene models are used to represent the multi-modal sensory information about the scene.Graph-based scene models are used to evaluate the diversity of scenes.The situation metrics computing network is developed to evaluate situation complexity metrics for scene models.We have used AD datasets such as nu Scenes,3D box detection algorithms such as SECOND,to verify the proposed situation analysis and evaluation framework.The experiment results show that the proposed methods can help evaluate multimodal AD data samples in terms of algorithm training and verification efficiency.Such a developed technology could be widely used for smart data collection in AD dataset construction and extreme scene construction in AD simulation.