Related Experiment Video
Updated: Mar 15, 2026
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
2.4K
Mask-Aware Spatiotemporal Classification of Millimeter-Wave Radar Point Cloud Sequences Using DGCNN and Transformer
1State Key Laboratory of Advanced Optical Communication Systems and Networks, Institute of Quantum Sensing and Information Processing, School of Sensing Science and Engineering, Shanghai Jiao Tong University, Shanghai 201100, China.
Sensors (Basel, Switzerland)
|March 14, 2026
Summary
This study introduces a novel framework for millimeter-wave radar point cloud classification, achieving 97.8% accuracy in identifying children, cats, and dogs. The method enhances recognition in complex environments by integrating spatial masking and temporal modeling.
Area of Science:
- Computer Vision
- Sensor Data Processing
- Machine Learning
Background:
- Non-contact target recognition in enclosed spaces demands robust solutions.
- Millimeter-wave radar offers privacy-friendly, low-light, and occlusion-robust sensing.
- Challenges include sparse point clouds, noise, multipath interference, and temporal point number fluctuations, degrading existing classification methods.
Purpose of the Study:
- To develop a spatiotemporal joint classification framework for millimeter-wave radar point cloud sequences.
- To address challenges posed by sparse, noisy, and temporally dynamic point clouds in complex environments.
- To improve the reliability and accuracy of life form identification in confined spaces.
Main Methods:
- Introduced a point mask mechanism in the spatial dimension to suppress invalid points and enhance local geometric representation.
- Integrated attention-based time series modeling in the temporal dimension to capture cross-frame dynamic patterns.
- Developed a spatiotemporal joint classification framework for millimeter-wave point cloud sequences.
Main Results:
- Achieved 97.8% accuracy in a three-classification task (Child, Cat, Dog).
- Ablation analysis confirmed the significant contributions of the mask mechanism and time series modeling.
- Demonstrated robust recognition performance in complex environments.
Conclusions:
- The proposed framework offers a deployable and generalized solution for millimeter-wave point cloud classification.
- The spatiotemporal approach effectively handles sparse, noisy, and dynamic point cloud data.
- This method significantly enhances the identification of life forms in confined spaces using millimeter-wave radar.