In this study, we explore the problem of adaptive vehicle trajectory control for different risk levels. Firstly, we introduce a sliding window-based car-following scenario extraction method, propose a new alternative traffic conflict assessment metric, and build a comprehensive traffic scenario library. Secondly, based on deep reinforcement learning (RL), we design an adaptive car-following trajectory control algorithm, which is called Deep Adaptive Control , to cope with different traffic risk levels. Thirdly, we design five metrics in terms of safety, comfort, and energy consumption, and experimentally compare Deep Adaptive Control with human drivers and RL benchmarks. The experimental results show the superiority of Deep Adaptive Control compared to human drivers and existing RL methods, which can follow the preceding vehicle closely in low-risk situations to improve traffic efficiency, keep distance from the preceding vehicle in high-risk situations to improve safety and be optimal in terms of comfort and fuel consumption metrics.
With the growing prevalence of autonomous vehicles and the integration of intelligent and connected technologies, the demand for effective and reliable vehicle speed control algorithms has become increasingly critical. Traditional car-following models, which primarily focus on individual vehicle pairs, exhibit limitations in complex traffic environments. To this end, this paper proposes an enhanced state representation for the application of multi-agent reinforcement learning (MARL) in platoon-following scenarios. Specifically, the proposed representation, influenced by feature engineering techniques in time series prediction tasks, thoroughly accounts for the intricate relative relationships between different vehicles within a platoon and can offer a distinctive perspective on traffic conditions to help improve the performance of MARL models. Experimental results show that the proposed method demonstrates superior performance in platoon-following scenarios across key metrics such as the time gap, distance gap, and speed, even reducing the time gap by 63%, compared with traditional state representation methods. These enhancements represent a significant step forward in ensuring the safety, efficiency, and reliability of platoon-following models within the context of autonomous vehicles.
In intelligent transportation systems, one key challenge for managing ride-hailing services is the balancing of traffic supply and demand while meeting passenger needs within vehicle availability constraints. Accurate origin–destination (OD) demand predictions can empower platforms to execute timely reallocation of cruising vehicles and improve ride-sharing services. Nonetheless, the complexity of OD-based demand prediction arises from intricate spatiotemporal dependencies and a higher need for precision compared to zone-based predictions, which leads to many unprecedented OD pairs. To tackle this issue, we design a comprehensive set of 102 features, including travel demand, passenger count, travel volume, liveliness, weather, and cross features. We also introduce an enhanced conformer model, which is composed of a single conformer block that integrates feedforward layers, multihead self-attention mechanisms, and depth-wise separable convolution layers. To address the cold-start problem and manage large values, we design a specific algorithm for OD pairs lacking training data and apply a technique to handle larger values. Our approach demonstrates a marked improvement in prediction performance, with an 18% decrease in the total travel demand error and up to a 47% reduction for certain larger values in some cases. Through extensive experiments on a dataset collected from a city, provided by a ride-hailing platform, our proposed methods significantly outperform the most advanced models.
10 June 2024 in IEEE Transactions on Systems, Man, and Cybernetics: Systems
Accurate visual perception and comprehensive scene understanding are critical for the safety and reliability of autonomous vehicles (AVs). Nevertheless, the efficacy of visual perception systems can be impaired by the intricacy of road scenes, and the existing scene understanding approach may be insufficient. Consequently, this study proposes an enhanced scene understanding model to achieve precise awareness of driving situations. Recognizing the limitations posed by the oversimplification of samples in current urban scene datasets, we selected critical frames from 336000 video frames, sourced from real-world driving environments, to assemble a more complex road scene (CRS) dataset. We integrated Residual Neural Network and pyramid scene parsing network architectures and refined them through class mapping and targeted network fine-tuning. Based on the segmentation outputs and the XGBoost algorithm, we identified the driving scenarios for the ego vehicle, enabling instantaneous driving situation analysis. The predictive model also evaluated the trajectory of interactive vehicles and estimated their kinematic states. Furthermore, we have conducted a thorough evaluation of scenario complexity, integrating the features described above. The findings indicate that our model achieves a segmentation accuracy of 78.8% in CRSs, with a twofold improvement in training efficiency. We also confirmed the effectiveness of the scene understanding approach through real-world road testing in China. This research provides insight into situation awareness within CRSs, thereby enhancing the visual perception capabilities of AVs. The implications of these results are substantial for their application in autonomous driving tests and advancing decision-making and control algorithms.