Generate Realistic Test Scenes for V2X Communication Systems

TL;DR

This paper introduces V2XGen, an automated tool for generating realistic test scenes to evaluate and improve V2X cooperative perception systems in autonomous driving, reducing manual effort and enhancing system robustness.

Contribution

V2XGen provides a high-fidelity, automated scene generation method with fitness-guided strategies, enabling efficient testing and retraining of V2X perception systems.

Findings

V2XGen can generate realistic and diverse test scenes.

Testing with V2XGen detects system errors effectively.

Retraining with generated scenes improves detection accuracy.

Abstract

Accurately perceiving complex driving environments is essential for ensuring the safe operation of autonomous vehicles. With the tremendous progress in deep learning and communication technologies, cooperative perception with Vehicle-to-Everything (V2X) technologies has emerged as a solution to overcome the limitations of single-agent perception systems in perceiving distant objects and occlusions. Despite the considerable advancements, V2X cooperative perception systems require thorough testing and continuous enhancement of system performance. Given that V2X driving scenes entail intricate communications with multiple vehicles across various geographic locations, creating V2X test scenes for these systems poses a significant challenge. Moreover, current testing methodologies rely on manual data collection and labeling, which are both time-consuming and costly. In this paper, we design and implement V2XGen, an automated testing generation tool for V2X cooperative perception systems. V2XGen utilizes a high-fidelity approach to generate realistic cooperative object instances and strategically place them within the background data in crucial positions. Furthermore, V2XGen adopts a fitness-guided V2X scene generation strategy for the transformed scene generation process and improves testing efficiency. We conduct experiments on V2XGen using multiple cooperative perception systems with different fusion schemes to assess its performance on various tasks. The experimental results demonstrate that V2XGen is capable of generating realistic test scenes and effectively detecting erroneous behaviors in different V2X-oriented driving conditions. Furthermore, the results validate that retraining systems under test with the generated scenes can enhance average detection precision while reducing occlusion and long-range perception errors.