Researchers Stefan Hägele, Adam Misik, and Eckehard Steinbach from the Technical University of Munich have developed a novel approach to enhance the capabilities of millimeter-wave (mmWave) radar systems for advanced perception tasks. Their work, titled “RadarFuseNet: Complex-Valued Attention-Based Fusion of IQ Time- and Frequency-Domain Radar Features for Classification Tasks,” was published in the IEEE Transactions on Pattern Analysis and Machine Intelligence.
Millimeter-wave radar has become a valuable tool for sensing in environments where traditional vision-based sensors struggle, such as detecting occluded objects or differentiating between various surface materials indoors. The non-linear nature of mmWave radar signals makes them well-suited for analysis using deep learning techniques. However, current methods for classifying occluded objects and materials based on in-phase and quadrature (IQ) data still have room for improvement.
The researchers proposed a bidirectional cross-attention fusion network that integrates IQ-signal and FFT-transformed radar features. This approach uses distinct complex-valued convolutional neural networks (CNNs) to process the data. By combining these features, the method achieves better performance and robustness compared to using standalone complex-valued CNNs.
In their experiments, the researchers demonstrated near-perfect material classification accuracy of 99.92% when samples were collected at the same sensor-to-surface distances used during training. Moreover, their method showed improved generalization ability, achieving an accuracy of 67.38% on samples measured at previously unseen distances. For occluded object classification, the accuracy improved from 91.99% using standalone complex-valued CNNs to 94.20% using the proposed approach.
For the energy sector, this research could have practical applications in enhancing the capabilities of radar systems used for monitoring and maintenance of infrastructure, such as power lines and pipelines. Improved material classification can aid in the detection of corrosion or other material degradation, while better occluded object detection can help identify potential hazards or obstructions in various environments. The enhanced performance and robustness of these radar systems can contribute to more efficient and reliable energy infrastructure management.
This article is based on research available at arXiv.