Traffic flow prediction is the foundation of urban traffic guidance and control, as well as the main function of an intelligent transportation system (ITS). Accurate traffic flow prediction is important for road users, traffic management departments, and private enterprises. However, traffic flows usually show a high degree of variability, correlation, and complex patterns in both temporal and spatial domain, which makes accurate traffic flow prediction a challenging task. How to capture the potential and dynamic spatial–temporal relationships of traffic data has been the bottleneck issue for intelligent transportation researchers. To solve the above problems, this article proposes an adaptive graph fusion convolutional recurrent network (AGFCRN) to model the temporal and spatial characteristics of traffic flow data dynamically and adaptively. An adaptive graph fusion convolution is proposed to discover the changing relationships between traffic volumes without a priori knowledge. It uses a self-learned node embedding to generate static graphs and combines current and historical states to generate dynamic graphs at each time step. A gated recurrent layer with residual structure is designed to mitigate the decay of prediction effects in long-term modeling. In addition, an attention layer incorporating self-learned node embedding is introduced in the AGFCRN to efficiently adjust the prediction pattern of each node. Experiments on several public data sets demonstrate that AGFCRN can achieve competitive performance compared to other typical and state-of-the-art methods.

Antenna design with traditional numerical methods is a highly computationally expensive task. In this letter, a complex-valued graph neural network (GNN) with residual connections is proposed to efficiently predict radiation pattern for antenna arrays with different geometric structures. The model uses GNN as backbone, and the topological structure and excitation of antenna array are encoded into graph representation. Complex weights and residual connections are designed to improve the accuracy of the model. Experiments are conducted on five kinds of arrays under amplitude-phase control. The experimental results show that the average mean absolute percentage error (MAPE) and the average root mean square error (RMSE) of the model on five types of antenna array are 1.86% and 0.032 respectively, while the time efficiency of the model is 6 orders of magnitude faster than that of traditional full-wave simulation.

Traffic flow prediction is the foundation of traffic scheduling and a major component of intelligent transportation systems (ITSs). Accurate traffic flow prediction is crucial for numerous real-world traffic applications. However, the complex and dynamic nature of spatiotemporal correlations in traffic flow data poses obstacles to this task. Another important factor hindering model improvement is the scarcity of spatio-temporal data. Meanwhile, most existing studies assume the traffic data collected by sensors in real scenarios is completely correct and trustworthy, thereby neglecting the anomalies and incorrect data. To address these issues, we propose the contrastive learning-based adaptive graph fusion convolution network with residual-enhanced decomposition strategy (CDAGF) for traffic flow forecasting, which performs a simple yet effective graph augmentation mechanism imposed only on the learned graphs and reserves the explicit graph to mitigate the data scarcity problem, and designs a negative filter to assign negative pairs based on timestamp information. Besides, CDAGF separates anomalous signals from valid traffic signals through a residual-enhanced decomposition strategy to weaken the impact of anomalies and further improve the prediction accuracy. Moreover, CDAGF generates static graph and dynamic graphs based on traffic signals as well as timestamp information and fuses the learned graphs with the explicit graph to model temporal and spatial characteristics dynamically and adaptively. The experimental results on five popular public data sets demonstrate that CDAGF achieves state-of-the-art performance with 18.29 mean absolute error (MAE) on the PEMS04 data set.

Prohibited items detection in X-ray images is an essential part of security inspection in various scenarios. Currently this work is mainly completed by expensive labor, as the detection accuracy of the automatic recognition algorithms needs to be further improved. In this paper, we propose the novel ideologies of whole-process feature fusion and local-global semantic dependency interaction to improve the automatic detection of prohibited items. Specifically, for whole-process feature fusion, we design Coordinated Fusion Backbone (CFB), Adaptively Refined Feature Pyramid (ARFP) and Selective Dense Feature Interaction (SDFI). For local-global semantic dependency interaction, we develop the ConvFormer (CF). CFB extracts and fuses the multi-scale pyramidal features of input image in parallel, bridging the gap between high-level semantic context features and low-level spatial-aware features. Then the local-global long-range dependencies are further modeled through CF to construct the initial fusion features that are fed into ARFP to be adaptively fused and refined to alleviate the semantic misalignment and information loss of pyramidal features, so as to produce the refined features focusing on identifiable regions. Finally, the SDFI performs in-depth fusion interactions on the refined features to obtain optimal feature representation. The above improvements are integrated into our model family PIXDet (T/S/L). Extensive experiments are conducted on the challenging SIXray, OPIXray, CLCXray and PIDray datasets, and experimental results show that our PIXDet detector family achieves remarkable detection results (91.2% mAP, 91.6% mAP, 60.4% AP and 80.0% AP by PIXDet-S) on these benchmarks with favorable competitiveness compared with other SOTA methods.

Visual odometry (VO) plays a vital role in Simultaneous Localization and Mapping (SLAM). Most of the current learning-based VO methods utilize Convolutional Neural Network (CNN) as framework. However, CNN is weak in integrating global context information. In the design of loss function, the majority of these methods ignore the restraint relationship between translation and rotation prediction. In this work we propose an end-to-end global-context-aware visual odometry scheme with Epipolar-Geometry-constrained loss function to estimate the relative 6-DoF poses of monocular camera. The proposed scheme designs an augmented-attention-enhanced global context block on top of contextual CNN to learn the long-range dependencies and internal correlation. To overcome the problem of mutual restraint between the translation and rotation errors, an Epipolar-Geometry-constrained loss function is developed to simultaneously improve the prediction accuracy of both translation and rotation. The evaluation results on public datasets and self-collected dataset show that the proposed system outperforms state-of-the-art learning-based VO methods with a large margin.