UP-TIFA: UWB Radar-Based People Counting via Time–Frequency Attention Neural Network

Xikang Jiang, Jiahang Guo, Chong Rao, Lin Zhang, Lei Li

Abstract: Ultrawideband (UWB) radar-based people counting (RPC) using deep learning-based methods has become a crucial technology for spatial awareness and monitoring in Internet of Things (IoT) applications. Recent deep learning approaches, particularly those combining convolutional neural networks (CNNs) with time-series processing modules such as transformers, have shown promise in RPC tasks. However, these modules were originally designed for applications like natural language processing (NLP). When directly applied to RPC, they may suffer from overfitting and low robustness due to the short-term and local correlation of radar signals. To address these challenges, an ultrawideband radar-based people counting method via time–frequency attention neural network is proposed, namely, UP-TIFA. UP-TIFA employs a dual-channel backbone network incorporating both time-domain and frequency-domain processing, enhancing spatial and temporal feature extraction. A time–frequency hybrid attention (TFHA) module is proposed, which integrates local attention mechanisms in both domains. A local sliding window restricts attention to spatially and temporally relevant regions, while a learnable gating mechanism adaptively fuses time and frequency domain outputs. To further improve model efficiency and generalization, a multihead orthogonal constraint (MOC) is introduced to enforce orthogonality among query and key projection matrices across different attention heads, reducing parameter redundancy. To handle with clutter from environmental noise and variations in electromagnetic wave attenuation due to distance and body orientation, an amplitude-phase joint optimization-based processing method is proposed, which enhances the signal-to-noise ratio (SNR) and stabilizes signal intensity across varying distances. A comprehensive radar dataset is collected in both open-hall and crowded indoor conference room environments for evaluation, featuring dynamic population counts ranging from 0 to 10 individuals in real-world conditions. Experimental results demonstrate that UP-TIFA achieves an average counting accuracy of 94.88%, outperforming the current state-of-the-art by 24.61%. Both the source code and the dataset are publicly available to facilitate further research.

External IDs:doi:10.1109/trs.2025.3605232