Human Activity Recognition
PulseAugur coverage of Human Activity Recognition — every cluster mentioning Human Activity Recognition across labs, papers, and developer communities, ranked by signal.
1 day(s) with sentiment data
LLM-based HAR frameworks will achieve widespread on-device deployment within 18 months.
The STELLA framework demonstrates that LLMs can be effectively utilized for on-device human activity recognition with real-time inference capabilities. As LLM efficiency improves and hardware capabilities increase, it's plausible that such frameworks will transition from research to practical, widespread deployment on consumer devices for personalized HAR applications.
Novel frameworks are actively addressing domain shift and data efficiency challenges in HAR.
Recent research on IBIS for Wi-Fi sensing and a new framework for inertial sensors indicates a strong industry focus on overcoming key limitations in HAR. These efforts aim to improve robustness against environmental changes and reduce the substantial data requirements for deep learning models, suggesting a trend towards more practical and efficient HAR solutions.
Federated learning approaches will become standard for privacy-preserving HAR on edge devices.
The exploration of FedAvg for HAR highlights the potential of federated learning to balance personalization and generalization while maintaining data privacy. As concerns over data privacy grow, federated learning offers a compelling solution for training HAR models on distributed edge devices without centralizing sensitive user data.
-
New CHEAR method boosts LLM efficiency for human activity recognition
Researchers have developed a new method called Confusion-Hierarchy Evidence-Aware Reasoning (CHEAR) to improve the efficiency and accuracy of human activity recognition using large language models (LLMs). CHEAR organize…
-
New framework uses ontology and personalized memory for human activity recognition
Researchers have developed OntoAR-HAR, a novel framework for human activity recognition (HAR) that leverages ontology-guided agentic retrieval-augmented generation (RAG) and personalized memory. This approach decomposes…
-
New MemFLoRA method slashes CNN memory use for edge adaptation
Researchers have developed a new parameter-efficient fine-tuning method called Memory-Floor LoRA (MemFLoRA) designed for adapting Convolutional Neural Networks (CNNs) at the edge. This method addresses the memory constr…
-
AI activity recognition models show performance gap for older adults
A new study published on arXiv highlights a significant performance gap in human activity recognition (HAR) models when applied to older adults, despite advancements in deep learning. Researchers found that models train…
-
New AI Studio Mimics Human Learning for Activity Recognition
Researchers have developed the Agentic Heuristic Learning (AHL) Studio, a novel approach to human activity recognition (HAR) that diverges from traditional gradient-based neural network training. Inspired by human cogni…
-
New framework enhances human activity recognition with virtual IMU data
Researchers have developed a new framework for improving human activity recognition (HAR) using wearable Inertial Measurement Units (IMUs). The proposed method addresses the challenge of limited labeled IMU data by inte…
-
New analysis reveals limited gains from individual domain generalization components in HAR
A new systematic analysis of domain generalization (DG) techniques for smartphone-based human activity recognition (HAR) reveals that individual components offer limited gains. The study, which involved over 410,000 exp…
-
HALO model advances IMU-based human activity recognition
Researchers have developed HALO, a novel foundation model for Human Activity Recognition (HAR) using inertial measurement units (IMUs). HALO addresses challenges of sensing heterogeneity and generalizing to unseen activ…
-
New TRI-HAR Framework Achieves Rotation-Invariant Activity Recognition
Researchers have developed TRI-HAR, a novel framework for Human Activity Recognition (HAR) using wearable IMUs. This system is designed to be robust to independent orientation shifts between IMUs at different body locat…
-
New OmniDecVAEs framework learns disentangled representations from multi-modal wearable data
Researchers have developed Omni-modal Variational Decomposition Autoencoders (OmniDecVAEs), a novel framework designed to learn comprehensive and disentangled representations from multi-modal wearable data. This system …
-
Wi-Fi sensing framework WiFuse boosts human activity recognition accuracy
Researchers have developed WiFuse, a novel framework for human activity recognition using Wi-Fi sensing. This dual-stream system fuses denoised time-domain amplitude variations with 2D-FFT-derived Delay-Doppler motion r…
-
DeepConvContext framework enhances human activity recognition with multi-scale temporal modeling
Researchers have developed DeepConvContext, a novel multi-scale framework designed to improve time series classification for Human Activity Recognition (HAR). This new approach addresses the limitations of traditional s…
-
New framework KineMIC enhances few-shot action synthesis for HAR
Researchers have developed KineMIC, a novel transfer learning framework designed to improve few-shot action synthesis for skeletal-based Human Activity Recognition (HAR). This method adapts text-to-motion diffusion mode…
-
Zero-Fi uses signal-language alignment for zero-shot Wi-Fi activity recognition
Researchers have developed Zero-Fi, a novel framework for Wi-Fi-based human activity recognition that utilizes contrastive signal-language alignment. This approach allows the system to recognize new activities without n…
-
RAG-HAR+ enhances LLM-based human activity recognition for edge devices
Researchers have developed RAG-HAR+, an extension of Retrieval-Augmented Generation for Human Activity Recognition (HAR). This new method optimizes cost-efficiency for LLM-based HAR on edge devices by strengthening retr…
-
Zero-Fi enables zero-shot Wi-Fi human activity recognition
Researchers have developed Zero-Fi, a novel framework for zero-shot Wi-Fi-based human activity recognition. This system utilizes contrastive signal-language alignment to learn unified representations from Wi-Fi signals …
-
New research tackles catastrophic forgetting in AI models · 7 sources tracked
Researchers are developing novel methods to address catastrophic forgetting in continual learning, a challenge where AI models lose previously acquired knowledge when learning new tasks. Several recent arXiv papers prop…
-
New JEPA framework enhances sensor-based activity recognition with unlabeled data
Researchers have developed a Joint Embedding Predictive Architecture (JEPA) framework to improve sensor-based human activity recognition. This new framework aims to learn robust representations from unlabeled datasets, …
-
New research offers data-efficient guidelines for inertial sensor deep learning
A new research paper proposes a data-efficient approach to deep learning for inertial sensor classification tasks. The study introduces a framework to estimate the minimum required training data size, finding that accur…
-
New theory tackles generalization in radar human activity recognition
Researchers have developed a theoretical framework to analyze generalization issues in through-the-wall radar (TWR) human activity recognition (HAR). The proposed framework establishes models for human kinematics, radar…