Depthwise Spatio-Temporal Heatmap Network for UWB Radar-Based Indoor Localization
UWB Radar · TinyML · TensorFlow Lite Micro · Depthwise Separable CNN · INT8 Quantization · ESP32-S3
- Developed a physics-guided TinyML system for device-free 2D localization of up to four people using raw CIR measurements from six UWB radars. Combined static amplitude, motion, breathing-sensitive and inter-antenna phase features.
- Designed a 22K-parameter fully convolutional network with temporal-sensor fusion, four depthwise separable residual blocks, and heatmap and person-count prediction heads.
- Achieved F1 of 0.9430, precision of 0.9068, recall of 0.9821 and mean localization error of 0.2129 m on locked test windows using the fully quantized INT8 model.
- Reduced the final TensorFlow Lite model to 42.2 KB with approximately 208.9 KB activation memory, within ESP32-S3 constraints of less than 800 KB model storage and 300 KB activation SRAM.