Deep learning has substantially advanced rail surface defect detection (RSDD). However, accurate defect detection under complex background interference and noise using existing methods remains difficult, while insufficient complementarity between models limits further performance gains. To address these challenges, a collaborative optimization network for RSDD, viz., CONet, is proposed herein. CONet uses two heterogeneous networks and a collaborative learning mechanism to exploit the complementary strengths of global spatial-domain semantics and local frequency-domain features. The first network is a frequency-guided diffusion perception network that enhances global semantic perception via adaptive filtering and matrix diffusion. The second is an adaptive state-space network based on the Vision Mamba backbone, which enables efficient cross-modal semantic integration via dynamic peer-adjacency fusion. The two networks collaboratively improve one another by exchanging intermediate encoded and prediction features. Specifically, a similarity-guided frequency-domain complementary strategy generates structure-aware feature representations, while cross-layer memory recall and attention-guided knowledge correction dynamically refine knowledge transfer during training. This collaborative design reduces background interference and strengthens defect-structure representation. Experiments on the industrial-grade RGB-D NEU RSDDS-AUG dataset show that CONet outperforms advanced comparison methods. Additional RGB-D SOD experiments support cross-task segmentation generalization, rather than direct cross-domain railway generalization. The deployment framework proposed herein remains prospective and requires future field validation. The code is available at https://github.com/xiaolili666/Collaborative-Optimization
Cooperative adaptive cruise control (CACC) must operate across platoons with heterogeneous vehicle responses and actuator delays, yet retraining a multi-agent reinforcement learning policy for every new configuration is interaction intensive. This paper develops Bellman-Residual Successor Feature (BRSF) transfer within the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) framework, yielding BRSF-MADDPG for extracting, selecting, and adapting transferable temporal knowledge across CACC tasks with a shared reward definition but different vehicle dynamics. BRSF first learns a shared transition representation that preserves reward-relevant information and information about state evolution, enabling source-domain successor features (SFs) to capture long-horizon behavior under heterogeneous dynamics. A small amount of target-domain data is then used to select the most compatible source actor–SF knowledge unit. During adaptation, the selected source-domain SF is kept fixed while an additive target-domain SF correction learns the target-specific temporal mismatch, avoiding direct overwriting of the selected source-domain SF. BRSF further decomposes the SF Bellman residual into reward-related and complementary components and calibrates their scales to balance their contributions to SF learning, knowledge selection, and target-domain SF correction. Theoretical analysis establishes fixed-policy SF contraction and convergence, explicit Bellman-residual bounds on SF and action-value errors, and guarantees for finite-probe knowledge selection and additive SF correction. Simulations demonstrate that BRSF-MADDPG achieves the lowest target-domain adaptation cost and the best converged reward among the compared transfer methods and ablated variants, while maintaining competitive CACC control performance.
Sudden accidents on freeways seriously disrupt traffic flow and easily induce secondary accidents, forming accident bottleneck segments with a tendency of congestion spreading. Inspired by the group operation rules of cellular automata, a centralized guidance strategy for autonomous vehicles (AVs) is designed to quickly release control commands and improve the safety and efficiency of mixed traffic flow. Firstly, the STCA model are improved, and a simulation environment that can characterize sudden accidents on freeways is built to generate traffic operation data. Secondly, a network mapping model from AVs operation rules to traffic flow state under accident conditions is established, and the approximation of which is obtained based on Koopman operator theory. Finally, the variable speed limit (VSL) and lane change (LC) guidance strategies of AVs are output based on data-driven model predictive control (DMPC), and the effectiveness of the proposed method is verified through comparative experiments. The simulation results show that the proposed generative data-driven MPC method can accelerate the congestion dissipation speed in different scenarios while ensuring that the generated data conforms to the characteristics of real traffic flow, which has the potential for practical application.
Accurate traffic state estimation is essential for transportation system optimization and management. Although vehicle trajectory data offer detailed spatio-temporal information for this purpose, their availability is usually limited to a small portion of vehicles, known as Connected Vehicles (CVs), due to the high deployment cost of large-scale sensing systems. Therefore, reconstructing complete trajectories for all vehicles from this partial data has become a critical prerequisite for precise traffic estimation. However, existing reconstruction methods are mainly designed for straight roads or single-lane intersections, thus struggling in complex scenarios due to: 1)incomplete representation of intricate road geometries, and 2)an inherent difficulty in balancing fine vehicle dynamics with macroscopic traffic flow rationality. Meanwhile, such methods are typically developed for specific traffic scenarios and therefore lack the capability of simultaneously modeling interrupted and non-interrupted flows. This paper proposes a scene-wide vehicle trajectory reconstruction method to address these issues, achieving unified modeling for both interrupted and non-interrupted flows. The proposed method adopts Cartesian coordinates and incorporates scene grids from graphic sensors as auxiliary inputs to represent complex geometries. A scene-to-scene, sequence-to-sequence auto-regressive architecture is constructed, centered on a scene-wide trajectory reconstruction network designed to collaboratively reconstruct trajectories while balancing microscopic and macroscopic dynamic rationality. Experiments conducted on the Intersection A scenario from the CitySim dataset and Peachtree dataset from NGSim demonstrate that the method achieves longitudinal MAEs of 2.52m–1.87m and RMSEs of 3.65m–3.24m at 5%–15% CV penetration rates for interrupted flow. The reconstructed trajectories also enable accurate estimation of queue length and travel delay. It also demonstrates high accuracy in lane-changing and conflict zone reconstruction, proving its effectiveness in complex geometric scenarios and its value as a foundation for high-fidelity traffic estimation.
In urban driving scenes, traffic participants exhibit highly non-uniform spatial distributions governed by road structures such as lanes and drivable areas. However, most existing 3D perception methods treat BEV features uniformly, limiting their ability to focus on interaction-critical regions. High-definition (HD) maps provide stable and structured road representations, offering a natural source of structural guidance for urban perception. Motivated by this observation, we propose MapDiffuser, a multi-modal perception framework that leverages HD maps as structural guidance to reshape sensor feature processing in BEV space for complex driving scenes. Specifically, Map-Conditioned Diffusion Guidance (MCDG) conditions the learning process during training using map information, guiding the model to capture the associations between road structure and sensor responses in the BEV feature space. Building upon this, Map-Aware Spatial Resampling (MASR) explicitly exploits map geometry during both training and inference to steer feature sampling and aggregation toward lane topology and drivable regions, enabling road structure to directly constrain the perception process at the spatial operation level. On the nuScenes benchmark, integrating MapDiffuser into the BEVFusion baseline yields overall gains of +1.4 NDS / +0.7 mAP. Moreover, MapDiffuser serves as a plug-and-play module for other BEV-based perception frameworks, consistently improving performance. We further evaluate its robustness under sensor-degraded conditions, where MapDiffuser achieves +1.5 NDS / +2.4 mAP in nighttime scenes and +0.8 NDS / +1.1 mAP under rainy conditions.
