Autonomous Driving Systems (ADS) are critical components of Intelligent Transportation Systems (ITS), where vehicle-level reliability has a direct bearing on road traffic safety. Evaluating ADS performance in complex environments remains challenging due to the absence of test oracles and the heavy reliance on deep learning. To address these challenges, this study proposes a novel metamorphic testing framework tailored for vision-based ADS. First, causal inference is employed to extract key environmental factors from high-dimensional observational traffic data, thereby reducing the test space. Second, a multi-objective optimization algorithm integrating causal counterfactual reasoning is developed to quantify the challenges associated with specific combinations of causal factors, enabling cost-effective exploration of test conditions. Third, low-risk source images are systematically transformed into hazardous driving scenes through a fine-tuned diffusion model, allowing ADS evaluation to be guided by metamorphic relations (MRs). Empirical experiments show that the proposed method achieves a higher fault detection ratio than the strongest baseline in four out of five ADS models, with relative gains ranging from 18.1% to 88.9%. Data augmentation experiments further demonstrate that incorporating MR-violating test cases can reduce ADS prediction errors by up to 13.67%, with these benefits preserved in real-world road traffic datasets through domain adaptation. This study highlights a new pathway for validating the reliability of vision-based ADS driven by deep learning, thereby supporting the deployment of safer road transportation. The source code for our methods and baselines is available athttps://github.com/SafeDL/AutoMetTest

Stereo matching achieves significant progress with iterative algorithms like RAFT-Stereo and IGEV-Stereo. However, these methods struggle in ill-posed regions with occlusions, textureless, or repetitive patterns, due to a lack of global context and geometric information for effective iterative refinement. To enable the existing iterative approaches to incorporate global context, we propose the Global Regulation and Excitation via Attention Tuning (GREAT) framework which encompasses three attention modules. Specifically, Spatial Attention (SA) captures the global context within the spatial dimension, Matching Attention (MA) extracts global context along epipolar lines, and Volume Attention (VA) works in conjunction with SA and MA to construct a more robust cost-volume excited by global context and geometric details. To verify the universality and effectiveness of this framework, we integrate it into several representative iterative stereo-matching methods and validate it through extensive experiments, collectively denoted as GREAT-Stereo. This framework demonstrates superior performance in challenging ill-posed regions. Applied to IGEV-Stereo, among all published methods, our GREAT-IGEV ranks first on the Scene Flow test set, KITTI 2015, and ETH3D leaderboards, and achieves second on the Middlebury benchmark. Code is available at https://github.com/JarvisLee0423/GREAT-Stereo.

Test the safety of Autonomous Driving Systems (ADS) with realistic traffic conditions is important to the insurance industry, legislators, and third-party technical services. Approaches for ADS testing can be divided into two main categories: physical test and virtual test, as shown in Fig. 1.

Visual-based object detection has become a crucial component in the realm of autonomous vehicles. However, conducting reliable testing for such systems remains unresolved. In this paper, we advocate for the application of causal inference to investigate the pivotal environmental factors influencing detection accuracy. Through the integration of diffusion models, we address the specialized conditional generalization of hazardous testing images. Our approach involves the construction of observational data to attribute key factors and fine-tune the diffusion model. Additionally, we introduce an optimal prompt words search method that strikes a balance between test coverage and level of challenge. Subsequently, leveraging these optimal prompts, we propose a cost-effective testing image generation through both "Text2Scene" and "Image2Scene" fashions. The experimental results indicate that, on the generalized dataset, the performance of object detection algorithms is the poorest, with the average detection accuracy decreasing from 0.81 to 0.285. Moreover, retraining object detection models on our generalized critical test cases can ultimately enhance algorithm performance, achieving a median accuracy improvement of up to 8.13%. Overall, our research proposes a novel approach to generalize test cases, thereby contributing to the advancement and deployment of safer autonomous vehicles.

Reinforcement learning has been shown to have great potential applications in autonomous driving. For collaborative driving scenarios, multi-agent reinforcement learning can be used to explore efficient and collaborative driving strategies. However, it still faces the challenge of non-stationary. Traditional methods focus on evaluating the similarities between the real state of the teammate and the modeled state. There is also the issue of partial observability. It can be addressed by establishing communication to share information with other surrounding agents. However, prior approaches overlook the efficient communication problem caused by unprocessed and redundant information. To tackle these two challenges, we propose an approach named Multi-Agent Collaboration via Cross-Attention and Communication (MACAC). MACAC leverages the agent’s local observations to analyze and capture environment and interaction information, while also incorporating teammate modeling through the exchange of concise state information via communication. In addition, to improve the learning process, we integrate the noisy advantage technique into MACAC to enhance the agent’s exploration capabilities. As a result, vehicles can effectively adapt to dynamic environments and exhibit efficient collaborative driving skills. In all, experiments conducted on an autonomous driving simulator demonstrate that our approach surpasses the performance of the baseline algorithms.

Automated driving systems (ADS) have made remarkable progress in recent years, yet their reliability and testability remain as significant challenges. The environmental conditions that ADS face are highly complex and may result in the disruption of autonomous vehicles. In this study, we propose an approach that leverages causal inference theory to analyze the impact of causal factors on automated driving visual modules. Our method uncovers the root key factors that affect visual perception performance. We further establish a Challenging Index to quantitatively characterize the causal effects of the key factors on perception failures. This quantitative index is subsequently utilized to generate risky scenarios. Through extensive experiments on various state-of-the-art automated driving visual algorithms, we demonstrate the effectiveness of the challenge index in evaluating the level of hazard in the deployment environment. Additionally, the proposed “challenge index guided search” method improves test efficiency by up to 8.95 times compared to the baselines while maintaining a balance between coverage diversity and the hazardous level of test scenarios. Our research offers a new perspective for analyzing and evaluating the impact of key factors on visual perception. This contributes to the reduction of test space and efficiency of the generation of high-value test scenarios, ultimately advancing the deployment of safer automated vehicles.