Scene Searching Based Robustness Assessment Of Deep Autonomous Driving Perception

In recent years,deep learning techniques applied to computer vision have made a breakthrough and then they are widely used in the field of autonomous driving.Including Apollo,Autoware,etc.many autonomous driving systems use deep learning techniques such as object detection,tracking,self-pose prediction,etc.The use of deep learning makes the system be more adaptable to complex environments and tasks.At the same time,it makes the autonomous driving system finally move from laboratory to reality and from an assist system to a fully autonomous system.However,the endless safety issues of autonomous driving systems have raised concerns.The deep learning system itself is highly nonlinear and uninterpretable,which provides difficulties in the safety analysis of the system.A large number of works related to the safety assessment of deep learning based autonomous driving systems have been proposed,including neuron coverage,scene coverage,input space partitioning,fault injection detection,and test dataset generation.However,they have limitations and cannot simultaneously combine the three points of understandability,flexibility,and ability to apply to highly nonlinear systems.In this paper,focusing on the understandability and flexibility of deep autonomous driving system assessment,we design and develop a deep autonomous driving assessment system based on scene search using the real-time rendering engine Unreal Engine and the CARLA autonomous driving test platform.The work in this paper includes the following points.1.Designed and implemented a assessment system for a deep autonomous driving system,and assessed the state-of-the-art deep conditional autonomous driving system,CILRS,which was able to search 520 scenes that caused abnormal output of the autonomous driving system from 1000 scenes.2.Designed a scene description method in real-time rendering engine,which can describe the relevance of objects in the scene more flexibly than the existing methods,and can be suitable for scene transformation search tasks,while being scalable.3.An efficient scene search algorithm for nonlinear systems is designed,which is able to perform a fault scene search task for an initial scene in an average time of 16.86 s.4.Designed a targeted road system for the deep conditioned autonomous driving system and built three representative assessment environments for the road system using the Unreal Engine editor.5.An analyzer was designed for the searched failure scenarios,which can analyze and determine the objects or weather conditions causing the failures.The analyzer in this paper determines from a statistical point of view that the areas where the CILRS system is prone to cause failures are on both sides of the road,rainy days and red or yellow objects are more likely to cause failures of this autonomous driving system.