Injection Simulation: an Efficient Validation Framework for Autonomous Driving System

Building a simulation environment is of great significance for realizing efficient and low-cost autonomous driving tests. However, even expensive 3D scene rendering methods cannot guarantee the authenticity of the data generated in the simulation system. Therefore, we expect to develop a simulation platform driven by real scene data, not only to pursue the aesthetic or realism of visual effects, but also to pay more attention to the generation of simulation data needed to verify the effectiveness of autonomous driving systems. In this paper, we propose a simulation framework for autonomous vehicles, which is called 'injection simulation'. The innovations of the framework include: i) A joint scene generation method based on high-precision semantic map and SUMO is proposed to generate various complex traffic scenes or construct 'corner cases' required for rapid testing; ii) A virtual bridge between ROS and SUMO is constructed to enable traffic flow vehicles and autonomous vehicles to interact in a full-duplex mode in the simulation environment to achieve the injection simulation of the fusion layer; iii) A real-time LiDAR data generation method based on a large-scale real point cloud map and virtual traffic flow is proposed to realize the injection simulation of the perception layer. The proposed framework can effectively test autonomous driving algorithms, and is easy to implement different test modes such as Software-in-the loop (SiL) and Vehicle-in-the-loop (ViL).