Jing Sun

For modern agricultural management, the accuracy of plant identification is crucial. However, the task becomes challenging because crops and weeds at early growth stages often exhibit similar color, leaf morphology, and texture in two-dimensional images captured under field conditions, despite their clear biological differences in terms of botanical species, root systems, and phenological characteristics. Furthermore, computing hardware in the field also has strict limits. Therefore, we developed the RT-BMTR network to handle these physical constraints. Within this architecture, image data is processed through a bilateral hybrid backbone named Bi-HMB. The DSFM captures small local details, and MambaVision understands the broader background information. Then, these features are fused by RepNCSPELAN4. We adopted this structure to reduce redundant calculations. Next, the model determines its bounding boxes using the Inner-ShapeIoU loss function. This geometric constraint improves the detection of small targets. When evaluated on the CropAndWeed dataset, our model achieved an average precision (AP) at IoU threshold 0.5 (AP⁵⁰) of 68.1%, AP⁷⁵ of 54.8%, and a mean AP averaged over IoU thresholds from 0.5 to 0.95 (AP^50–95) of 50.9%. Detection precision recorded 26.5% for small objects, 44.7% for moderate ones, and with 59.3% for large objects. Rates for the first two categories saw enhancements of 16.2% and 4.6%. Overall, our modified model outperforms the original RT-DETR baseline. We also shrank the overall parameter count by 30.1%, alongside a 4.2% decrease in computational demand.

The safe operation of metro systems, particularly in densely populated cities, relies on the effective management of crowd gatherings within stations. Existing research primarily focuses on camera-based crowd counting but fails to consider crowd movement across multiple interconnected spaces, limiting its effectiveness in complex metro environments. This paper proposes a real-time Edge-Cloud collaborative video analytics system for Crowd Gathering Event Detection (EC-CGED), integrating an edge-cloud collaboration mechanism, spatial knowledge model, and an event-driven adaptive dynamic scheduling strategy. The system enables fine-grained, real-time monitoring of crowd dynamics and provides early warnings of potential crowd gatherings. At the edge nodes, real-time video streams are decoded and analyzed to extract crowd counting indicators for various station areas and transit channels. These indicators, along with spatial knowledge of the metro station layout, are aggregated at the central cloud server to enhance the accuracy of crowd gathering event detection. Additionally, we introduce an event-driven adaptive dynamic task scheduling strategy to optimize computational resource allocation, improving system efficiency. By enabling timely detection and proactive alerts for crowd gatherings, the EC-CGED system enhances metro station safety and operational efficiency, addressing critical challenges in urban public transportation management.

Existing 3D human pose estimation methods use fixed computation strategies processing diverse action sequences, leading to computational redundancy for simple actions, insufficient high-frequency information capture for complex actions, and low long-sequence processing efficiency. To address these issues, this paper proposes a Spatial–Temporal Complexity-Aware Adaptive Computation Framework (CAAPoseFormer). First, a spatial–temporal coupled complexity quantification module is built to integrate spatial dispersion and temporal motion variance for graded action complexity quantification. On this basis, a time–frequency dual-domain adaptive pruning strategy is proposed to dynamically allocate temporal window length and frequency-domain DCT coefficients on demand. Furthermore, a mask-guided sparse interaction encoding mechanism is designed to enable efficient parallel computation of variable-length features by shielding invalid padding regions. Experiments on the Human3.6M dataset show that, versus the baseline PoseFormerV2, the proposed method cuts parameters by 85.3% and computational cost by 64.8% while retaining comparable accuracy (MPJPE 44.2 mm), boosting unit computational efficiency 2.8×. Moreover, compared with state-of-the-art (SOTA) methods like MHFormer and MotionBERT, our method reduces computational costs (MACs) by 97.4% and nearly three orders of magnitude, respectively. This framework effectively breaks the inference bottleneck of high-precision models on low-power hardware, suiting latency-sensitive real-time applications well.

Revealing the topographic gradient differentiation of land use in mountainous cities is essential for optimizing land-use allocation in complex terrain regions. Taking a typical mountainous urban area Nan’an District of Chongqing as a case study, this paper examines the topographic gradient characteristics and evolutionary patterns of its land-use structure using a Topographic Position Index and a topographic distribution index(DI), based on multi-temporal Sentinel-2 imagery and DEM data. The results indicate that: (1)Land use in Nan’an District is dominated by construction land and forest land; construction land is concentrated in low topographic gradient areas, whereas forest land predominates in medium and high gradient zones. (2)Topographic gradient differentiation of land use has progressively intensified: in low-gradient areas, construction land has expanded while cultivated land has shrunk, and in medium- and high-gradient areas, forest land has maintained a stable dominance. (3)The distribution advantages and structural characteristics of land use exhibit marked spatio-temporal variation across different topographic gradient levels. (4)The correlation between topography and land-use distribution is significant and persistent,which confirms that topography is a core constraining factor shaping and driving land-use patterns and their evolution in mountainous cities.

The Contrastive Language--Image Pre-training (CLIP) model, pre-trained on large-scale web image-text pairs, exhibits remarkable generalization capabilities and demonstrates exceptional `zero-shot' performance across various image tasks. Extending the advantages of CLIP pre-training to the video domain has remained a prominent research challenge. Previous methods attempt to bridge the video-image gap through either introducing additional fine-tuned components or freezing CLIP parameters during fine-tuning, aiming to maintain both strong supervised learning performance and CLIP's generalization ability. However, achieving an optimal balance between supervised performance and generalizability proves challenging. We propose a CLIP-Based Dual Temporal Decoupling Network for Video Action Recognition (TDN-CLIP), which preserves supervised capabilities while enhancing generalization. Our framework employs an Instance-aware Temporal Decoupling Module (ITDM) to decouple CLIP-extracted frame-level features into content-based features and style-residual features. The style-residual features are then processed by a Dual Temporal Global-Local Attention (DTA) mechanism to capture temporal dynamics: the Global Temporal Attention (GTA) models long-range dependencies across frames, while the Local Temporal Attention (LTA) captures fine-grained correlations between adjacent frames. Through cascaded global-local temporal collaborative attention, we achieve feature recalibration. The enhanced style-residual features are subsequently fused with content-based features to form domain-invariant action semantics, producing discriminative video representations that improve model generalizability. The effectiveness of our method is validated through zero-shot, Base-to-Novel, and few-shot experiments conducted on four public benchmark datasets.

The objective of visible-infrared person re-identification is to accurately match pedestrian images captured in different modalities. Since these images are taken from varying viewpoints by different cameras, the cross-modal detection task must address both modality discrepancies and camera variations. Many existing approaches primarily focus on minimizing inter-modality differences to enhance retrieval accuracy, often overlooking the impact of camera viewpoint differences. To tackle these challenges, this article introduces a hierarchical feature decoupling network. First, the network decouples and extracts camera-related and camera-irrelated features separately to mitigate the effects of camera variations. Second, it addresses modality differences by extracting modality-independent features. Additionally, an adversarial decoupling loss is employed to further disentangle identity-irrelevant information from identity-relevant features, thereby boosting the system’s accuracy and robustness. Extensive experiments conducted on the SYSU-MM01 and RegDB datasets validate the effectiveness of the proposed method.

Urban Functional Zones (UFZs) are spatial units of the city divided according to specific functional activities. Detailed identification of UFZs is vital for optimizing urban management, guiding planning and design, and promoting sustainable development. However, existing UFZ recognition methods face significant challenges, such as difficulties in effectively integrating multi-source heterogeneous data, capturing dynamic spatiotemporal patterns, and addressing the complex interrelationships among various data types. These issues significantly limit the applicability of UFZ mapping in complex urban scenarios. To address these challenges, this paper proposes a tripartite neural network (TriNet) for multimodal data processing, including Remote Sensing (RS) images, Point of Interest (POI) data, and Origin–Destination (OD) data, fully utilizing the complementarity of different data types. TriNet comprises three specialized branches: ImgNet for spatial features extraction from images, POINet for functional density distribution features extraction from POI data, and TrajNet for spatiotemporal pattern features extraction from OD data. Finally, the method deeply fuses these features through a feature fusion module, which utilizes a two-layer fully connected network for deep fusion, allowing the model to fully utilize the interdependencies among the data types, significantly improving the UFZ classification accuracy. The experimental data are generated by mapping OpenStreetMap (OSM) vector into conceptual representations, integrating images with social sensing data to create a comprehensive UFZ classification benchmark. The method achieved an overall accuracy of 84.13% on the test set of Chongqing’s main urban area, demonstrating high accuracy and robustness in UFZ classification tasks. The experimental results show that the TriNet model performs effectively in UFZ classification.