To guarantee rational decision-making and safety of intelligent transportation systems and autonomous driving, existing vehicle trajectory prediction (VTP) methods extract spatial and temporal features from complex traffic environments to achieve accurate forecast results. However, they generally do not use the frequency-domain information inherently embedded in vehicle trajectory data, resulting in lower prediction accuracy. To solve this problem, we propose a joint spatio-temporal-frequency-aware feature fusion (STFA-FF) method for VTP. Firstly, an intention recognition network is proposed to integrate spatial features and temporal features to infer driving intentions with high precision. Secondly, to fully utilize frequency-domain features, we present a multi-scale frequency-domain feature extraction (MSFDFE) module to map vehicle trajectory data into the frequency domain, incorporate the high-frequency attenuation mask to suppress high-frequency noise, and deeply integrate short-term variations with long-term trends. Additionally, a frequency-domain channel selection (FDCS) module is proposed to dynamically select key frequency channels related to driving modes. Furthermore, a multi-domain feature fusion prediction network is proposed to process the spatial, temporal and frequency-domain features to generate the final trajectory prediction results. Finally, experimental results demonstrate that the proposed method significantly outperforms mainstream approaches in prediction accuracy and robustness.

With the rise of low-altitude economy applications, 6G communication systems place stricter demands on optical fiber amplifiers, requiring wider bandwidth, higher gain, and better spectral uniformity. Backbone networks for low-altitude integrated sensing and communication systems, in particular, call for high-performance amplification to support robust data transmission and reliable sensing. However, traditional multi-pump Raman fiber amplifiers (RFAs) are no longer adequate for meeting the performance demands of distributed fiber optic sensing networks in such scenarios. To address this problem, this paper proposes a machine learning-enhanced multi-pump RFA for high-performance optical backbone in low-altitude sensing and communication. The back propagation neural network (BPNN) is employed to accurately model the nonlinear relationship between pump parameters and amplification performance, facilitating adaptive and fine-grained control over signal gain, which is critical for maintaining stable and efficient data transmission across dynamic and heterogeneous communication scenarios. Moreover, the artificial bee colony (ABC) algorithm is integrated to perform global optimization of pump wavelengths and power configurations, thereby improving overall system bandwidth, gain characteristics, and operational robustness under diverse and unpredictable network conditions. The experimental results demonstrate that the proposed method achieves superior prediction accuracy, enhanced stability, and greater adaptability compared to conventional algorithms.

With the rapid growth of the low-altitude economy, emerging applications such as urban air mobility and smart logistics demand reliable and low-latency beamforming for unmanned aerial vehicle-to-vehicle (UAV-to-UAV, U2U) communications in millimeter-wave (mmWave) bands under highly dynamic and non-line-of-sight (NLOS) conditions. Traditional beam alignment methods relying on exhaustive search or channel feedback incur heavy training overhead and degraded accuracy in rapidly varying environments. To address these challenges, we propose multi-modal perception-assisted fusion network (MPFusionNet), a multi-modal perception-enhanced Transformer framework for predictive beamforming. Our approach leverages heterogeneous onboard sensing data including global positioning system (GPS), red-green-blue (RGB) cameras, LiDAR, and radar altimeters, incorporates a dynamic time warping (DTW)-based alignment mechanism, and embeds geometry-aware priors within a perceiver input-output (PerceiverIO)-based fusion architecture to achieve robust spatiotemporal representation. Experiments on a simulated U2U dataset show that MPFusionNet attains a top-3 beam prediction accuracy of 97.59%, substantially surpassing conventional models. These results demonstrate the effectiveness of multi-modal learning in improving robustness and generalization of predictive beamforming for future autonomous aerial communication systems.

Vehicle trajectory prediction (VTP) is an important task that can enhance the safety and efficiency of autonomous driving. However, existing VTP methods often struggle to fully extract spatiotemporal features, resulting in inaccurate prediction results. To solve this problem, we propose a hierarchical long short-term memory and graph attention network (HLSTM-GAT) model. First, we design a hierarchical network architecture to model different spatiotemporal features. The first-layer network (FLN) focuses on short-term trajectory information and immediate vehicle interactions to generate preliminary candidate trajectories. In the FLN, long short-term memory (LSTM) encoder processes historical trajectories to extract temporal features of vehicles, while a graph attention network (GAT) handles the LSTM-encoded outputs to capture spatial features between vehicles. Second, the second-layer network (SLN) combines the candidate trajectories generated by the FLN with historical trajectories to form comprehensive trajectories for accurate prediction. Subsequently, SLN adopts a GAT to process these comprehensive trajectories to precisely model the spatial relationships between vehicles. Moreover, we propose two distance threshold-based dynamic GAT models to perfectly capture spatial features between vehicles. They construct vehicle spatial interaction relationships at two distinct levels based on real-time distances in the FLN and predicted future distances in the SLN, respectively. These dynamic GAT models effectively filter out irrelevant information and enable the proposed method to consider potential future interactions. Finally, we conduct extensive experiments on two publicly available datasets, and experimental results demonstrate that the proposed method outperforms other state-of-the-art VTP methods in terms of prediction accuracy, robustness and computational efficiency.

Traffic flow prediction (TFP) plays a crucial role in optimizing road resource allocation and alleviating traffic congestion. However, existing TFP methods have limitations in capturing the complex spatiotemporal dependencies from traffic data, resulting in low prediction accuracy. To solve this problem, we propose a hybrid long short-term memory (LSTM) and graph attention network (GAT) model based on multi-scale temporal features (MSTF-LG) to predict traffic flow. Firstly, we employ trigonometric functions (TF) to process timestamp information to extract its periodicity and continuity features. These features are then integrated with traffic data to construct a comprehensive input representation. We further extract recent traffic data as well as daily, weekly, and monthly periodic data from this representation. Secondly, we adopt the LSTM encoder to process recent traffic data to extract recent trend features, and apply LSTM encoders to handle daily, weekly and monthly periodic data to extract periodic features. Furthermore, we employ the GAT to process these LSTM-encoded multi-scale temporal features of traffic data to capture dynamic spatial characteristics. The normalized GAT outputs are fed into the LSTM decoder to effectively capture the dynamic temporal changes of traffic data, and then a linear layer transforms the output of the LSTM decoder into TFP results. Finally, experimental results demonstrate that the proposed method outperforms existing TFP methods in terms of prediction accuracy, robustness and computational efficiency.

Vehicle position prediction (VPP) is of great significance for navigation planning and traffic safety of intelligent vehicles. In general, particle filtering (PF) uses global navigation satellite system (GNSS) to implement VPP. However, it does not consider geographic layer information (GLI) and its particle weight is not combined with the real-world geographic position information, which leads to insufficient prediction preparation. To resolve this problem, we propose a novel PF-based VPP method by using three-dimensional convolutional neural network and long short-term memory (3D CNN-LSTM) network model. First, for data preprocessing, we extract kinematic information features from GNSS, and evenly divide the area around each GNSS point into multiple grids and calculate the probability of grids center belonging to each GLI type. In addition, in order to better reflect the relationship between two consecutive positions due to the factors such as the conversion angle, we construct tilted cells to represent possible positions of each vehicle at any time. Second, a novel 3D CNN-LSTM model is designed to calculate the vehicle occurrence probability (VOP) in each tilted cell by processing the GLI and GNSS data, which can optimize the PF weight of each particle, and then improve PF to make more precise position prediction. Finally, the experimental results demonstrate that the proposed VPP method can improve the cell prediction accuracy, and then significantly improve the position prediction precision.

This correspondence discusses recognition of traffic modes and behavior patterns based on Global Navigation Satellite System (GNSS) data. The traffic modes (e.g., walk, car, train, etc.) are firstly inferred, and then their behavior patterns (e.g., left-turn, right-turn, turn-around, etc.) are further identified. Because both traffic modes and behavior patterns are strongly influenced by geographical circumstances, their recognitions are enhanced by geographical layer information (e.g., building, road, water, etc.). At one specific GNSS point, its surrounding area is uniformly sliced as grids, and the probabilities for grid centers belonging to six different geographical layers are calculated based on whether these centers lie inside or outside of the minimum rectangles containing polygons indicating different geographical objects. Finally, the six-dimensional probability matrix is processed and compressed as a geographical information vector by the convolutional neural network (CNN). The latter is then combined with kinematic metrics such as velocity, acceleration, and moving direction from GNSS data, and serially input into a long short-term memory (LSTM) network to predict traffic modes and behavior patterns. Experimental results validate that the geographical information does enhance the performances of two recognition tasks. The CNN+LSTM framework retains the powers of CNN and LSTM, and outperforms classical machine learning algorithms.