human-activity-recognition

This paper proposes pretraining the CABiGRU temporal architecture on diffusion-generated sensor windows, then fine-tuning on real DEO smartwatch data, to improve recognition of underrepresented eating/drinking activities, achieving 90.6% balanced accuracy. The approach shows that diffusion-based synthetic data better preserves multivariate temporal structure than SMOTE/GAN augmentation for handling class imbalance in human activity recognition.

This paper evaluates Jev, a fixed general-purpose probabilistic decision model, for training-free human activity recognition on accelerometer data, finding it far below supervised models (macro-F1 0.038–0.118 vs 0.686–0.907) and showing that adding more numerical sensor features actually degrades performance unless paired with deterministic semantic renderings.

This paper introduces a wireless smart insole platform and machine-learning workflow for recognizing human activities like sitting, standing, walking, and unstable walking to support continuous monitoring in elderly care, using pressure and inertial signals with high accuracy.

This paper proposes a coverage-aware virtual IMU augmentation framework to address data scarcity in human activity recognition, generating and selecting synthetic sensor data to improve model performance with limited labeled data.

The paper presents AHL Studio, a tool using agentic heuristic learning to create executable, inspectable policies for human activity recognition without traditional training, targeting edge deployment.

A research paper evaluating CHAP, a CNN-BiLSTM model, for classifying sedentary behavior from wearable accelerometer data, focusing on transfer from hip to wrist placement and the benefits of finetuning.

RAG-HAR+ is a retrieval-first, cost-optimized extension of RAG-HAR for human activity recognition from wearable sensors. It uses a retrieval designer agent and majority voting to reduce LLM usage while maintaining accuracy, and demonstrates feasibility for edge deployment.

Inertia-1 is a fully open exploration of wearable motion foundation models using over 18.2 million hours of accelerometer data, systematically studying data, model, and training choices across 15 datasets for tasks like activity recognition and disease prediction.

STELLA is an efficient sensor-to-LLM translation framework for on-device human activity recognition that compresses multi-channel inertial sensor data into compact latent tokens, achieving new state-of-the-art performance across seven public HAR datasets while supporting on-device personalization and real-time inference.

This paper systematically compares seven state-of-the-art sensor fusion methods for multi-modal human activity recognition on the HARMES dataset, showing that Gated Multi-modal Fusion achieves the highest macro F1-score of 0.82, outperforming the baseline by 6 percentage points.

This paper presents WISE-HAR, an ensemble deep learning framework for WiFi-based human activity recognition, achieving robust performance and generalization across scenarios with minimal accuracy drops.

This paper systematically explores hybrid KAN and MLP architectures for IMU-based human activity recognition, achieving a 5.33% average macro F1 improvement over pure MLP baselines.