To address the issues of insufficient samples, limited scene diversity, missing perspectives, and low resolution in existing UAV-based pedestrian detection datasets, this paper proposes a novel UAV-based pedestrian detection benchmark dataset named the Novel Surveillance View (NSV). This dataset encompasses diverse scenes and pedestrian information captured from multiple perspectives, and introduces an innovative data mining approach that leverages tracking and optical flow information. This approach significantly improves data acquisition efficiency while ensuring annotation quality. Furthermore, an improved pedestrian detection method is proposed to overcome the performance degradation caused by significant perspective changes in top-down UAV views. Firstly, the View-Agnostic Decomposition (VAD) module decouples features into perspective-dependent and perspective-independent branches to enhance the model’s generalization ability to perspective variations. Secondly, the Deformable Conv-BN-SiLU (DCBS) module dynamically adjusts the receptive field shape to better adapt to the geometric deformations of pedestrians. Finally, the Context-Aware Pyramid Spatial Attention (CPSA) module integrates multi-scale features with attention mechanisms to address the challenge of drastic target scale variations. The experimental results demonstrate that the proposed method improves the mean Average Precision (mAP) by 9% on the NSV dataset, thereby validating that the approach effectively enhances pedestrian detection accuracy from UAV perspectives by optimizing perspective features.
Multimodal entity linking (MEL) grounds entity mentions in text-image contexts to entries in a structured knowledge base; however, most existing systems still decompose a multimodal document into independent mention-level decisions. This formulation overlooks a central tension of real-world MEL: the evidence needed to disambiguate an ambiguous mention is often distributed across co-occurring mentions, visual context, and cross-modal consistency, rather than being contained in the mention itself. To address this limitation, we propose Collaborative Entity Linking through Multichannel Interaction (CELMI), a four-channel framework that jointly models textual semantics, visual perception, cross-modal alignment, and inter-mention collaboration. CELMI employs dual-level textual alignment, text-guided visual gating, learned cross-modal projection, and attention-based entity graph propagation. To stabilize joint optimization, we further introduce a multi-channel consistency objective that combines per-channel contrastive losses with an overall ranking loss, reducing channel dominance and representation collapse. Among conventional non-LLM/VLM MEL models, CELMI achieves the strongest MRR and Hits@1 performance on WikiMEL, RichpediaMEL, and WikiDiverse, with Hits@1 scores of 89.03% on WikiMEL and 83.02% on RichpediaMEL; LLM/VLM systems remain stronger on WikiDiverse, positioning CELMI as a lightweight complement. Progressive stress ablation shows a 38.50 percentage-point absolute Hits@1 drop when the architecture is reduced to a single unimodal endpoint, confirming that the gains arise from synergistic channel interaction rather than isolated module effects.
Accurate traffic flow forecasting is a fundamental yet challenging task in intelligent transportation systems, because of the highly nonlinear and complex nature of traffic dynamics. This paper presents a novel Spatio-Temporal Graph Neural Network with Integrated Semantics and Topology (STIGNN). Unlike existing approaches, STIGNN jointly exploits node-level semantic features and graph topological dependencies to enhance spatial representation learning. Specifically, spatial correlations are captured through a ChebNet–GAT hybrid mechanism, which combines multi-hop structural information with semantic attention. Temporal dependencies are modeled using a Gated Temporal Convolutional Network (Gated TCN), enabling efficient learning of long-range sequences. Furthermore, the dual-attention design over node semantics and network topology allows STIGNN to be directly applied to inductive prediction tasks, ensuring strong generalization across diverse network settings. Extensive experiments on two real-world benchmark datasets, METR-LA and PEMS-BAY, demonstrate that STIGNN consistently outperforms state-of-the-art baselines.
1 November 2026 in IEEE Transactions on Neural Networks and Learning Systems
Designing effective graph neural networks (GNNs) for diverse tasks requires substantial manual effort, especially when dealing with the intricate interplay between topological structures and semantic information in graph data. Existing automated approaches often rely on off-the-shelf model combinations that remain superficial, resulting in generated model frameworks that are hardly interpretable and limited by the capabilities of the combined models. To address this gap, we propose an evolutionary-based framework that systematically searches for optimal GNN architectures while adaptively combining topological and semantic insights. This approach reduces human intervention, accommodates a wide range of graph tasks, and ensures robust performance in complex scenarios. Extensive experiments on real-world datasets demonstrate that our method outperforms both handcrafted GNNs and mainstream AutoML models in terms of accuracy, offering a scalable and adaptive solution for GNN design.