Robust Real-Time Human Activity Recognition via 2D Tensor CNN With Change Point Detection and Confidence-Based Prediction Filtering
Abstrak
Falls are a serious health issue, particularly among the elderly, where delayed detection can lead to severe injuries or even death. Conventional fall detection systems based on cameras or wearable devices suffer from limitations in terms of privacy, comfort, and reliability. To overcome these challenges, this study proposes a real-time Human Activity Recognition (HAR) system using mmWave FMCW radar integrated with a Convolutional Neural Network (CNN), combined with Confidence Based Prediction Filtering (CBPF) and Change Point Detection (CPD) algorithms. Radar point cloud data are transformed into two-dimensional representations, namely Doppler Range, Doppler Time, and Range Time, which are then classified into four main activities: sitting, standing, walking, and falling. The current CNN-based HAR radar approach still faces limitations in computational efficiency and prediction stability because classification is performed on each frame without considering significant temporal changes in activity. In this paper, CBPF algorithm is designed to improve classification stability by applying a confidence threshold and enforcing temporal consistency (consecutive prediction count) before updating the final label. Meanwhile, the CPD module detects significant changes in Doppler patterns that indicate transitions in activity and only forwards radar frames that show such significant changes to the CNN for classification, thereby improving computational efficiency. In addition, a stride windowing parameter analysis was conducted to support real-time processing with minimal delay while maintaining high accuracy. Experimental results show that integrating CBPF and CPD significantly enhances real-time classification accuracy, increasing accuracy from 75% (without CBPF and CPD) to 93.33%, while maintaining a maximum system latency of 0.83 s, which is suitable for real-time deployment. The proposed system was successfully implemented in a real-time graphical user interface (GUI) displaying point cloud, heatmap, and stable classification outputs. These findings demonstrate that the proposed approach effectively delivers accurate and reliable activity detection, making it highly applicable for real-world scenarios such as elderly monitoring and smart healthcare systems.
Identifier
- DOI: https://doi.org/10.1109/access.2026.3696665
- OpenAlex: https://openalex.org/W7162325125