Automated test case generation for self-driving cars using CCTV videos

Official URL: https://risc01.sabanciuniv.edu/record=b2579393_(Table of contents)

Self-driving cars are more and more included in our daily lives with recent advancements in the fields of artificial intelligence and robotics. Self-driving cars are typically tested in simulation and real-life with various types of tests in different scenarios. To ensure the safety of self-driving cars we must be able to conduct realistic test cases in simulation and real life. The most realistic and highly convincing tests include test cases from real-life accidents. However, currently, these tests are generated manually with humans still in the loop. We introduce a generic and scalable way to realistically generate automated test cases from any car accident video that is available. To show the flexibility of our study we use various YouTube videos that we have no prior information about for evaluation. The proposed method consists of three steps, namely analysis, scene reconstruction, and test case generation. Our method is fully automated and an end-to-end solution to generate test cases. In the analysis step, we run the input videos through an image processing pipeline that consists of six internal steps. Then we reconstruct the crash in a 3D physics engine. Finally, we generate various test cases from a set of automatically pre-defined or user-defined parameters. The test cases can be used with any autonomous driving stack that satisfies the communication requirements through a simulation bridge. We evaluate our results with a user study and a set of case study experiments that are conducted on the popular open-source autonomous driving stack, Apollo.