With the advent of the commercialization phase for autonomous vehicles (AVs), the evaluation of their intelligence has become essential for regulators. However, existing evaluation methods are still largely traditional and experience-based. Therefore, this study proposed a subjective-objective mapping evaluation (SOME) method to evaluate the intelligence of high-level autonomous driving systems (ADSs). First, a five-dimensional evaluation metric system was developed to represent the overall performance of AVs during testing or actual driving. Next, the performance of AVs in a real-world driving dataset was evaluated based on the large language model (LLM). This approach could significantly enhance evaluation efficiency, achieving a passing rate of 86.75% in the Turing test. Finally, a deep neural network with an attention mechanism was trained using quantified metrics and an LLM-based evaluation label to serve as the evaluation model. Ablation experiments were conducted on both the LLM-based evaluation strategy and the evaluation model, demonstrating the necessity of each module. During the application phase, the evaluation model's effectiveness and reliability in scenario-based testing were validated using data from the OnSite Autonomous Driving Challenge results, and comparative analysis was conducted against traditional evaluation methods. The results show that our model's evaluations align closely with those of human experts and outperform traditional evaluation methods.

With the rapid advancement of autonomous driving and Intelligent Transportation Systems (ITS), roadside perception—an essential component of vehicle-to-everything (V2X) communication—has become a critical foundation for large-scale traffic monitoring and data-driven safety decisions. However, under adverse environmental conditions such as rain, snow, fog, and nighttime, current roadside perception systems often face limited training data, high annotation costs, and significant performance degradation due to poor model robustness. To tackle these challenges without relying on additional labeled data, this paper introduces ROSE (Roadside Oversight-guided Scenario Enhancement), a unified closed-loop framework designed to enhance the resilience and adaptability of roadside multi-modal perception systems. ROSE integrates three key components: (1) RISA, a first-order physics-guided cross-modal augmentation module that generates physically plausible and semantically aligned adverse-weather samples; (2) SSL Coupling, a cross-modal self-supervised learning network that facilitates robust feature alignment; and (3) a curriculum scheduling mechanism guided by vision-language models (VLM) to adaptively prioritize learning difficulty. Experimental results show that ROSE achieves competitive detection accuracy while providing improved robustness and cross-modal consistency across adverse weather scenarios, demonstrating strong generalization potential for roadside multi-modal perception under challenging environmental conditions. These findings suggest that ROSE offers a practical and effective pathway toward building resilient V2X cooperative perception systems.