Yunong Tian

Object detection in real-world settings frequently suffers from low-visibility challenges, such as backlighting, rain, snow, and occlusion as well as large variations in object scale. To address these issues, we present a novel Multi-scale Enhanced Cross-Modal Adaptive Fusion Network for RGB-T object detection, called CMAF-Net, which improves cross-modal fusion, multi-scale contextual modeling, detection head design, and localization loss optimization. In this framework, an Adaptive Gated Fusion Module (AGFM) is introduced at the mid-level fusion stage to strengthen the information exchange and complementary representation between the visible and infrared branches. In addition, a Multi-scale Contextual Attention Fusion module (MCAF) is used to strengthen multi-scale contextual modeling. An Adaptive Multi-scale Fusion Head (AMFH) is also designed to adaptively aggregate features across different scales, thereby improving detection performance for objects of varying sizes. During training, a Dynamic Weighted Intersection-over-Union loss (DW-IoU) is introduced to optimize bounding box regression, improving localization stability and regression accuracy for samples of varying quality through dynamic focusing and center-distance constraints. To validate the model’s performance, systematic experiments were conducted on three datasets: M3FD, LLVIP and VEDAI. The results demonstrate that CMAF-Net exhibits superior performance compared to other state-of-the-art object detection methods while maintaining a good balance between detection accuracy and computational cost. The experimental results also indicate that CMAF-Net can preserve robust detection capability in difficult scenarios, including rainy, foggy, and cluttered environments, exhibiting good environmental adaptability and robustness.

Fire poses a significant threat to human life and property safety, making early fire detection and warning crucial. However, conventional systems primarily rely on temperature and smoke sensors, which exhibit limited detection range, resulting in significant monitoring blind spots and a high risk of missed alarms. In contrast, video surveillance-based fire detection methods can achieve large-scale, blind-spot-free real-time monitoring. Although deep learning breakthroughs have spurred recent progress in early fire detection research, building a universal and robust detection system remains a critical challenge. To address these issues, this paper proposes a Dual-Branch Feature Fusion Network (DBFFNet) based on YOLOv11 for accurate smoke and fire detection in fire scenes. First, we introduce a Dual-Branch Feature Fusion (DBFF) structure within the feature extraction component to enhance the model’s ability to extract and utilize feature information. Further, we propose an R-C3k2 module inspired by GELAN and optimized through RepVGG reparameterization. This module serves as an improved building block, providing richer feature representations and reducing inference time, thereby boosting the mAP. Additionally, we incorporate the SCSA attention mechanism within a C2PSA module, which effectively combines channel and spatial attention to fully leverage multi-semantic information and further improve recognition accuracy. To achieve lightweight deployment, we introduce the ADown module as an improved downsampling convolution, reducing model parameters through a lightweight design. We further introduce the MPDIoU loss function specifically to enhance the detection accuracy of small flame and smoke targets in early fire scenarios. Comprehensive experiments show that in fire target detection tasks, DBFFNet achieves an mAP50 of up to 0.919 and an mAP50-95 of up to 0.718, surpassing the previous SOTA model, YOLOv8-EMSC, which achieved an mAP50 of 0.871 and an mAP50-95 of 0.615, by absolute improvements of 0.048 and 0.103 respectively, and establishing a new SOTA performance.

Against the backdrop of digital transformation, this study constructs an analytical framework for the influence mechanism of the data-driven dynamic capabilities of foreign trade SMEs from the perspective of digital intelligence immunity, aiming to clarify the complex relationships among influencing factors and multi-combination paths for capability improvement. The research employs the fuzzy AHP-DEMATEL method to quantify the complex influence relationships among factors and uses fsQCA to analyze the configuration paths of high-level data-driven dynamic capabilities. Results show that digital intelligence management and analysis, digital intelligence supervision and early warning, and digital intelligence ecosystem are key drivers of data-driven dynamic capabilities, with digital intelligence talents serving as a guarantee and digital foundation as a foundation. The study identifies the following two core paths for forming high-level capabilities: “management–talent–ecology collaboration” and “early warning–technology–mechanism enhancement.” It concludes that foreign trade SMEs should strengthen digital intelligence management and ecological construction, improve early warning mechanisms, and adopt multi-pronged approaches to build data-driven dynamic capabilities, providing a theoretical basis for their digital transformation and capability upgrading.

With the widespread adoption of mobile devices and the increasing availability of user trajectory data, accurately predicting the next location a user will visit has become a pivotal task in location-based services. Despite recent progress, existing methods often fail to effectively disentangle the diverse and entangled behavioral signals, such as collaborative user preferences, global transition mobility patterns, and geographical influences, embedded in user trajectories. To address these challenges, we propose a novel framework named Multi-Perspective Hypergraphs with Contrastive Learning (MPHCL), which explicitly captures and disentangles user preferences from three complementary perspectives. Specifically, MPHCL constructs a global transition flow graph and two specialized hypergraphs: a collective preference hypergraph to model collaborative check-in behavior and a geospatial-context hypergraph to reflect geographical proximity relationships. A unified hypergraph representation learning network is developed to preserve semantic independence across views through a dual propagation mechanism. Furthermore, we introduce a cross-view contrastive learning strategy that aligns multi-perspective embeddings by maximizing agreement between corresponding user and location representations across views while enhancing discriminability through negative sampling. Extensive experiments conducted on two real-world datasets demonstrate that MPHCL consistently outperforms state-of-the-art baselines. These results validate the effectiveness of our multi-perspective learning paradigm for next-location prediction.

The unlawful use of noncooperative drones, or unmanned aerial vehicles (UAVs), poses serious threats to public safety and societal security, necessitating robust monitoring solutions. However, the detection of drones, particularly those flying remotely, is often hindered by the limited accuracy in identifying small targets. Additionally, challenges related to insufficient lightweight design and complexities in practical deployment further exacerbate the issue. To address these challenges, we propose a quad-head network, QHNet, designed to provide scalable, adaptable, and efficient drone detection across diverse scenarios. QHNet is available in five scalable model sizes—Nano (N), Small (S), Medium (M), Large (L), and Extra Large (X)—ensuring real-time detection tailored to varying resource constraints and operational requirements. The core of QHNet lies in its innovative quad detection head (QDH), which introduces an additional detection layer to perform secondary feature extraction for small targets, enhancing detection precision. Furthermore, the architecture integrates adaptive spatial feature fusion to improve scale invariance and overall detection accuracy. To further optimize performance, QHNet incorporates a four-scale feature fusion network (FSF) and a specialized small target IoU loss function (STIoU) to refine detection precision. Simultaneously, the lightweight coarse-to-fine processing unit (LCF) and SCDown downsampling (SCD) form a comprehensive lightweight scheme, significantly reducing computational overhead while maintaining high detection efficacy. Extensive experiments were conducted on DUT-Plus, an augmented version of the DUT Anti-UAV dataset enhanced through data augmentation. The experimental results confirm that QHNet attains outstanding performance, providing an advanced balance between detection accuracy and computational efficiency across all model sizes.

LiDAR-based loop closure detection is a crucial part of realizing robust SLAM algorithms for intelligent vehicles with LiDAR sensors. Existing methods often reduce the keypoint dimension to encode the global descriptor, which sacrifices the freedom of loop detection and correction. Based on the 6-DOF rigid transformation property of spatial triangles, we propose an algorithm for extracting and describing 3D keypoints from high-resolution spinning LiDAR intensity images to encode triangle descriptors, termed intensity triangle descriptor (ITD) . In comparison to the direct extraction of keypoints from the point cloud, the use of image-derived feature points provides additional photometric texture information and better handles uneven spatial density of the point cloud, which is advantageous in unstructured and geometrically degraded scenes. To enhance the stability of keypoints, the spatial positions of multi-frame image feature points are registered to a keyframe by an odometer for voxel downsampling and non-maximum suppression, with the objective of reducing unstable feature points. For high discrimination, the neighbor image patches of each vertex (keypoint) are aggregated to estimate a Gaussian mixture model (GMM) as the keypoint signature. An efficient two-stage loop closure detection method is then proposed for ITD, consisting of candidate retrieval based on triangle side lengths and vertex GMMs, followed by geometric verification of matched descriptor pairs. The effectiveness of the proposed method is evaluated on the STheReO, FusionPortable, and our self-collected datasets.

Zero-shot learning (ZSL) enables models to recognize categories not encountered during training, which is crucial for categories with limited data. Existing methods overlook efficient temporal modeling in multimodal data. This paper proposes a Temporal–Semantic Aligning and Reasoning Transformer (TSART) for spatio-temporal modeling. TSART uses the pre-trained SeLaVi network to extract audio and visual features and explores the semantic information of these modalities through audio and visual encoders. It incorporates a temporal information reasoning module to enhance the capture of temporal features in audio, and a cross-modal reasoning module to effectively integrate audio and visual information, establishing a robust joint embedding representation. Our experimental results validate the effectiveness of this approach, demonstrating outstanding Generalized Zero-Shot Learning (GZSL) performance on the UCF101 Generalized Zero-Shot Learning (UCF-GZSL), VGGSound-GZSL, and ActivityNet-GZSL datasets, with notable improvements in the Harmonic Mean (HM) evaluation. These results indicate that TSART has great potential in handling complex spatio-temporal information and multimodal fusion.