Boting Qu

Falls in the older population constitute a significant public health issue, motivating the development of automated fall risk assessment methods. However, existing approaches based on wearable sensors or depth cameras are often limited by hardware cost and deployment complexity. To this end, we present DualFRA, a vision-based deep learning approach that leverages standard cameras and integrates complementary mobility information from Timed Up and Go (TUG) and Five-Times Sit-to-Stand (FTSTS) test videos for comprehensive fall risk assessment. DualFRA first applies 3D pose estimation to extract skeletal sequences from the two clinical tests. For each sequence, multiple complementary representations are constructed, including a spatio-temporal skeletal graph and auxiliary views that emphasize joint configurations and motion dynamics. These representations are then processed by a spatio-temporal feature encoder combining self-attention-enhanced graph convolution and multi-scale temporal modeling to capture discriminative movement patterns. The resulting dual-stream features are integrated through cross-attention to produce a coherent semantic embedding for fall risk prediction. Experimental results averaged across multiple independent runs demonstrate that DualFRA attains accuracy and weighted-F ₁ scores of 87.90% and 87.83%, surpassing baselines with average performance gains of 2.92% and 2.97%, respectively. Ablation analyses further confirm the benefits of the dual-stream design.

Hiking activities originated during the European Industrial Revolution (Frison-Roche & Jouty, 1996) and developed into a globally popular ecological recreational activity (Moira et al., 2021). Nati...

Traffic flow prediction is a critical task in Intelligent Transportation Systems (ITS). However, the complex spatiotemporal dynamics, periodicity, and anomalous variations in traffic data pose significant modeling challenges. To address this, this paper proposes a novel Wavelet Residual and Multi-Scale Dynamic Graph Convolution Network (WR-MDGCN). It first applies the discrete wavelet transform (DWT) to decompose traffic signals into low-frequency and high-frequency components, and then performs layer-wise residual refinement to progressively extract spatiotemporally consistent periodic and anomalous patterns, laying a foundation for differentiated modeling. We then construct signal-specific dynamic graphs to enable targeted modeling of fine-grained spatial dependencies. Based on these graphs, WR-MDGCN employs multi-scale dynamic graph convolution to capture both stable patterns and abrupt spatial changes. Periodic signals are modeled via dynamic graph convolution, while anomalous signals are captured using hierarchical diffusion convolution to enhance sensitivity to sudden changes. To adapt to the distinct characteristics of periodic and anomalous signals, a differentiated global information processing mechanism is designed. For periodic signals, the global memory mechanism strengthens long-term dependencies, while for anomalous signals, the dynamic global pooling mechanism considers global dynamic feedback. Experiments on six real-world datasets demonstrate that WR-MDGCN consistently outperforms baselines in both accuracy and stability.

Accurate traffic flow prediction is a key application in Internet of Things (IoT)-enabled intelligent transportation systems (ITS). However, complex spatiotemporal dependencies and non-stationary anomalies make reliable forecasting particularly challenging. To address these issues, we propose Bi-STDGCN, a Bidirectional Spatiotemporal Diffusion Graph Convolutional Network with Multi-Scale Anomaly-Guided Attention. The core idea is to jointly model anomaly-aware temporal variations and bidirectional spatial diffusion within a unified multi-source dynamic graph framework. First, an Anomaly-Guided Feature Enhancement Module (AGFEM) employs a difference-driven attention mechanism guided by multi-scale temporal patterns derived from wavelet-based decomposition, enhancing the model’s ability to perceive periodic disruptions and sudden traffic fluctuations. Second, Bi-STDGCN captures directional dependencies in both time and space by integrating a bidirectional diffusion graph convolutional recurrent unit (BDDGCRU) for temporal modeling to reduce prediction lag under anomalies, and a bidirectional diffusion GCN (BD-GCN) to model upstream propagation and downstream feedback patterns in the spatial domain. Third, the Multi-source Dynamic Attention Graph Generator (DAGG) incorporates static topology, temporal attributes, real-time traffic states, and hidden node states to construct robust time-varying spatial graphs. Extensive experiments on five real-world datasets demonstrate that Bi-STDGCN achieves state-of-the-art accuracy and robustness, while ablation and visualization analyses further validate the effectiveness of each component.

Falls among the elderly and other at-risk populations pose a significant public health challenge, necessitating innovative methods to detect and intervene early. While automated fall risk assessment methods using wearable sensors or depth cameras have been proposed, the physical and psychological burden of wearing extra sensors, and the high cost of specialized equipment like Kinect, limits their practical adoption. To tackle this challenge, this paper presents a novel machine learning-based fall risk assessment approach called FRAVM, which operates on Five times Sit-To-Stand (FSTS) test videos captured with standard, widely available cameras to identify individuals requiring fall prevention interventions. To enhance the practicality of FRAVM, 3D pose estimation is applied to generate vision-rich 3D body keypoints, mitigating the challenges posed by restricted camera angles. Median-average filtering is used to reduce noise caused by video shaking and pose estimation inaccuracies, while a new Dynamic Time Warping (DTW)-based matching algorithm is designed to handle interference from irrelevant individuals appearing in the video. Furthermore, a novel Attention-augmented Spatial-Temporal Graph Convolutional Network (AST-GCN) is developed for reliably identifying the action in each frame, enabling accurate computation of key kinematic features for fall risk prediction. Experimental evaluations on a dataset of 450 FSTS test videos demonstrate the high performance of FRAVM, with detection accuracy, F1 score, and ROC-AUC achieving 88.64%, 88.58%, and 98.86%, respectively. The ablation analysis confirms that the components in FRAVM are essential for achieving optimal results.

Falls among elderly individuals and other at-risk populations represent a significant public health concern, underscoring the need for proactive and accessible fall risk assessment methods. Although prior systems using wearable sensors or depth cameras have demonstrated potential, their widespread adoption is often constrained by user discomfort, setup complexity, and the high cost of specialized equipment such as Kinect. To address these limitations, we propose GTAE-FRA, a novel vision-based deep learning framework for fall risk assessment that leverages standard, widely available RGB cameras to record Five times Sit-To-Stand (FSTS) test videos. GTAE-FRA begins with a robust video processing pipeline that includes 3D pose estimation, noise filtering, and Dynamic Time Warping (DTW)-based matching to extract reliable Body Skeleton Key Point (BSKP) sequences. These sequences are then fed into GTrans, an attention-augmented encoder combining message passing aggregation mechanism with Transformer-based spatial-temporal modeling for nuanced feature extraction. To overcome limited data availability, six skeleton-oriented data augmentation strategies are applied, which significantly enhance the diversity and robustness of training samples. Experimental results on a dataset of 450 FSTS videos demonstrate the effectiveness of GTAE-FRA, achieving impressive performance with detection accuracy, weighted F1 score, macro F1 score, and AUC of 87.00%, 87.01%, 88.92%, and 97.96%, respectively. These results represent average improvements of 2.49%, 2.66%, 2.24% and 1.78% over baseline methods across the respective metrics.