Perception Data Generation via Synthesized 3D Scenes and Closed-Loop Optimization in Autonomous Driving Simulation

Simulation is crucial for developing autonomous driving systems, where AI algorithms process sensor data to perceive environments. However, training these algorithms require vast, diverse data, and existing methods face limitations. Manual scene construction in simulation platforms is slow and lacks scalability, while rendering-based approaches often produce inconsistent image sequences, failing to support high-quality algorithm training. Additionally, such methods cannot synthesize multi-modal data essential for perception models. To alleviate these problems, this paper proposes an automated system to generate multi-modal perception data (image, point cloud, vehicle status) by synthesizing diverse 3D scenes within a simulation platform. By directly constructing 3D scene models, the system eliminates perspective inconsistencies inherent in rendering-based methods. Experimental validation showing the system generates consistent, high-quality data for autonomous driving.