Selected work

Representative Papers

Impact Detection in Fall Events: Leveraging Spatio-Temporal Graph Convolutional Networks and Recurrent Neural Networks Using 3D Skeletons Data

Jul 28, 2026

This study addresses the challenge of distinguishing genuine fall impacts from non-impact balance losses in fall monitoring. To this end, the authors propose a spatiotemporal graph modeling approach based on 3D skeletal data. By constructing a spatiotemporal graph of human joints and leveraging a Spatial-Temporal Graph Convolutional Network (STGCN) to extract spatial-temporal features, the method further integrates GRU and BiLSTM modules to enhance temporal dynamics modeling for precise identification of fall impact moments. This work presents the first integration of STGCN with bidirectional recurrent neural networks for fall detection, achieving over 90% accuracy on an enhanced version of the UP-Fall dataset. The proposed approach significantly improves the discrimination between true and false falls, and the refined dataset is publicly released to support future research in this domain.

0 citationsRead paper

FLASH: Efficient Impact Fall Detection with Unified Hypergraph State-Space Model

Jul 28, 2026

This work addresses the limitations of existing skeleton-based fall detection methods, which struggle to accurately model the multi-joint coordination patterns at the moment of impact and suffer from high computational costs due to complex temporal modeling, hindering real-time deployment. To overcome these challenges, we propose a novel approach that integrates biomechanically informed hypergraph structures with the Mamba selective state space model. Our method leverages a single-matrix hypergraph guided by biomechanical priors and employs an adaptive feedback mechanism to efficiently capture both joint coordination and temporal dynamics. By uniquely combining biomechanics-driven hyperedges with a linear-complexity state space model, the proposed framework achieves state-of-the-art accuracy on the UP-Fall and UMAFall datasets while maintaining real-time inference capability, low computational overhead, and strong zero-shot cross-dataset generalization performance.

0 citationsRead paper

DistillH-Mamba: A Hypergraph-Mamba-Based Knowledge Distillation Model for Efficient Impact Fall Detection

Jul 03, 2026

Accurate identification of the precise moment of impact during elderly falls remains challenging, as existing methods often involve high computational complexity and are difficult to deploy in real-time scenarios. This work proposes DistillH-Mamba, a novel architecture that uniquely integrates hypergraph neural networks with the Mamba state space model to effectively capture high-order inter-joint relationships, long-range temporal dependencies, and abrupt motion changes. To further enhance efficiency, the authors introduce a relational knowledge distillation strategy that preserves critical spatiotemporal structural information while significantly compressing the model. Evaluated on the UP-Fall and UMAFall datasets, the proposed method achieves an impact detection accuracy of 97.38% and reduces inference time by 73.8% compared to the teacher model, substantially outperforming current state-of-the-art approaches.

0 citationsRead paper

dziribot: rag based intelligent conversational agent for algerian arabic dialect

Feb 02, 2026

This work addresses the scarcity of high-quality linguistic resources for Algerian Arabic (Darja)—a low-resource dialect hindered by non-standardized spelling, frequent code-switching with French, and dual-script usage—which impedes the development of intelligent dialogue systems. To overcome these challenges, we propose a hybrid conversational agent architecture tailored for Darja, integrating dedicated natural language understanding with retrieval-augmented generation (RAG) to support structured service workflows and dynamic responses grounded in enterprise knowledge bases. We evaluate three approaches: sparse features, traditional machine learning, and a fine-tuned Transformer model (DziriBERT). Notably, this study presents the first deployment of a dialect-level RAG dialogue system in a real-world business setting. The fine-tuned DziriBERT achieves state-of-the-art performance in intent recognition, significantly outperforming baseline models and demonstrating robustness and scalability in handling spelling variations and rare intents.

0 citationsRead paper

GZSL-MoE: Apprentissage G{é}n{é}ralis{é} Z{é}ro-Shot bas{é} sur le M{é}lange d'Experts pour la Segmentation S{é}mantique de Nuages de Points 3DAppliqu{é} {à} un Jeu de Donn{é}es d'Environnement de Collaboration Humain-Robot

Sep 23, 2025

To address the generalized zero-shot learning (GZSL) challenge in 3D point cloud semantic segmentation—where training data lacks unseen classes and adaptation to dynamic human-robot collaboration scenarios is difficult—this paper proposes a novel method integrating generative modeling with a Mixture of Experts (MoE) mechanism. For the first time, MoE architectures are embedded into both the generator and discriminator of Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) to jointly model feature distributions of seen and unseen classes. Additionally, pre-trained KPConv features are leveraged to enhance point cloud representation robustness. Evaluated on the COVERED dataset, the method achieves significant improvements in both seen and unseen class segmentation performance, outperforming state-of-the-art approaches in H-score and harmonic mean metrics. These results validate its effectiveness and generalization capability for recognizing previously unknown objects in real-world collaborative settings.

0 citationsRead paper

Recent publications

Latest Papers

Impact Detection in Fall Events: Leveraging Spatio-Temporal Graph Convolutional Networks and Recurrent Neural Networks Using 3D Skeletons Data

Jul 28, 2026

This study addresses the challenge of distinguishing genuine fall impacts from non-impact balance losses in fall monitoring. To this end, the authors propose a spatiotemporal graph modeling approach based on 3D skeletal data. By constructing a spatiotemporal graph of human joints and leveraging a Spatial-Temporal Graph Convolutional Network (STGCN) to extract spatial-temporal features, the method further integrates GRU and BiLSTM modules to enhance temporal dynamics modeling for precise identification of fall impact moments. This work presents the first integration of STGCN with bidirectional recurrent neural networks for fall detection, achieving over 90% accuracy on an enhanced version of the UP-Fall dataset. The proposed approach significantly improves the discrimination between true and false falls, and the refined dataset is publicly released to support future research in this domain.

0 citationsRead paper

FLASH: Efficient Impact Fall Detection with Unified Hypergraph State-Space Model

Jul 28, 2026

This work addresses the limitations of existing skeleton-based fall detection methods, which struggle to accurately model the multi-joint coordination patterns at the moment of impact and suffer from high computational costs due to complex temporal modeling, hindering real-time deployment. To overcome these challenges, we propose a novel approach that integrates biomechanically informed hypergraph structures with the Mamba selective state space model. Our method leverages a single-matrix hypergraph guided by biomechanical priors and employs an adaptive feedback mechanism to efficiently capture both joint coordination and temporal dynamics. By uniquely combining biomechanics-driven hyperedges with a linear-complexity state space model, the proposed framework achieves state-of-the-art accuracy on the UP-Fall and UMAFall datasets while maintaining real-time inference capability, low computational overhead, and strong zero-shot cross-dataset generalization performance.

0 citationsRead paper

DistillH-Mamba: A Hypergraph-Mamba-Based Knowledge Distillation Model for Efficient Impact Fall Detection

Jul 03, 2026

Accurate identification of the precise moment of impact during elderly falls remains challenging, as existing methods often involve high computational complexity and are difficult to deploy in real-time scenarios. This work proposes DistillH-Mamba, a novel architecture that uniquely integrates hypergraph neural networks with the Mamba state space model to effectively capture high-order inter-joint relationships, long-range temporal dependencies, and abrupt motion changes. To further enhance efficiency, the authors introduce a relational knowledge distillation strategy that preserves critical spatiotemporal structural information while significantly compressing the model. Evaluated on the UP-Fall and UMAFall datasets, the proposed method achieves an impact detection accuracy of 97.38% and reduces inference time by 73.8% compared to the teacher model, substantially outperforming current state-of-the-art approaches.

0 citationsRead paper

dziribot: rag based intelligent conversational agent for algerian arabic dialect

Feb 02, 2026

This work addresses the scarcity of high-quality linguistic resources for Algerian Arabic (Darja)—a low-resource dialect hindered by non-standardized spelling, frequent code-switching with French, and dual-script usage—which impedes the development of intelligent dialogue systems. To overcome these challenges, we propose a hybrid conversational agent architecture tailored for Darja, integrating dedicated natural language understanding with retrieval-augmented generation (RAG) to support structured service workflows and dynamic responses grounded in enterprise knowledge bases. We evaluate three approaches: sparse features, traditional machine learning, and a fine-tuned Transformer model (DziriBERT). Notably, this study presents the first deployment of a dialect-level RAG dialogue system in a real-world business setting. The fine-tuned DziriBERT achieves state-of-the-art performance in intent recognition, significantly outperforming baseline models and demonstrating robustness and scalability in handling spelling variations and rare intents.

0 citationsRead paper

GZSL-MoE: Apprentissage G{é}n{é}ralis{é} Z{é}ro-Shot bas{é} sur le M{é}lange d'Experts pour la Segmentation S{é}mantique de Nuages de Points 3DAppliqu{é} {à} un Jeu de Donn{é}es d'Environnement de Collaboration Humain-Robot

Sep 23, 2025

To address the generalized zero-shot learning (GZSL) challenge in 3D point cloud semantic segmentation—where training data lacks unseen classes and adaptation to dynamic human-robot collaboration scenarios is difficult—this paper proposes a novel method integrating generative modeling with a Mixture of Experts (MoE) mechanism. For the first time, MoE architectures are embedded into both the generator and discriminator of Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) to jointly model feature distributions of seen and unseen classes. Additionally, pre-trained KPConv features are leveraged to enhance point cloud representation robustness. Evaluated on the COVERED dataset, the method achieves significant improvements in both seen and unseen class segmentation performance, outperforming state-of-the-art approaches in H-score and harmonic mean metrics. These results validate its effectiveness and generalization capability for recognizing previously unknown objects in real-world collaborative settings.

0 citationsRead paper