Friday 21 March 2025
A team of researchers has made a significant breakthrough in the field of human activity recognition using radar technology. By applying a novel approach that combines low-rank adaptation and serial-parallel adapter fine-tuning, they have developed a method that can accurately identify various human activities from radar signals.
The researchers used a dataset collected by the University of Glasgow, which features six specific activities: walking, sitting, standing, drinking, picking up an object from the floor, and falling. They employed a Vision Transformer (ViT) model, pre-trained on natural images, and fine-tuned it for use with radar-based Time-Doppler signatures.
The key innovation lies in the joint fine-tuning of both the weight space and feature space. The low-rank adaptation technique is used to transfer knowledge from the pre-trained model to the radar signals, while the serial-parallel adapter fine-tuning enhances the extraction of fine-grained features.
In experiments, the proposed method outperformed existing approaches, achieving an accuracy rate of 96.61% on the University of Glasgow dataset. Notably, it excelled in recognizing highly confusable activities such as drinking and picking up, which are often challenging to distinguish.
The implications of this work extend beyond the realm of human activity recognition. Radar technology has numerous applications in fields like healthcare, transportation, and security, where accurate identification of human actions can have significant benefits. For instance, in healthcare, radar-based fall detection systems could be used to monitor elderly individuals or patients with mobility issues.
This study demonstrates the potential of combining radar signals with cutting-edge computer vision techniques to develop robust and efficient systems for human activity recognition. The researchers’ innovative approach has opened up new avenues for exploring the capabilities of radar technology in various domains, paving the way for further advancements in this field.
Cite this article: “Radar-Based Human Activity Recognition Achieves High Accuracy with Novel Fine-Tuning Approach”, The Science Archive, 2025.
Radar, Human Activity Recognition, Computer Vision, Machine Learning, Deep Learning, Vit Model, Low-Rank Adaptation, Serial-Parallel Adapter Fine-Tuning, Time-Doppler Signatures, Fall Detection
