This paper proposes RobHAR, a millimeter wave radar-based human activity recognition (HAR) system mounted on a movable robot platform for healthcare monitoring. The key highlights are:
Sparse point cloud global embedding: A light-PointNet (LPN) backbone is used to extract global features from sparse and imbalanced point cloud data generated by the millimeter wave radar. A segment-wise point cloud augmentation (SPCA) algorithm is developed to enhance the quantity and quality of the training data.
Spatio-temporal HAR model: A bidirectional lightweight LSTM (BiLiLSTM) model is proposed to learn the temporal patterns from the time-distributed global point cloud features.
Transition optimization: A transition optimization strategy that integrates Hidden Markov Model (HMM) and Connectionist Temporal Classification (CTC) is introduced to improve the accuracy and robustness of continuous HAR.
Extensive experiments on three datasets demonstrate that the proposed RobHAR system significantly outperforms previous methods in both discrete and continuous HAR tasks, while achieving high efficiency for real-time deployment on a robot-mounted edge computing platform.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Zhanzhong Gu... at arxiv.org 05-06-2024
https://arxiv.org/pdf/2405.01882.pdfDeeper Inquiries