UWB Impulse Radar for People Counting With Convolutional Neural Network on Microcontrollers

We present an exploration of the advancements in ultrawideband (UWB) radar technology applied to people counting in the Internet of Things (IoT) applications. We introduce a novel lightweight convolutional neural network (CNN) model specifically designed for implementation on microcontrollers with low power consumption. Our proposed model successfully overcomes the limitations of existing methods, achieving an impressive accuracy rate of 99.38% for counting to ten people while randomly walking in a small area of
$5 \times 5$
m. In addition, its lightweight architecture enables effortless integration with resource-constrained microcontrollers, enabling efficient execution of intelligent IoT tasks. Notably, our method maintains the accuracy even after quantization, with the model retaining 98.22% accuracy while reducing its size by more than half. Benchmarking results demonstrate the model’s efficiency, with inference times of less than 48 ms on a wide range of STM32 microcontrollers. Furthermore, the model with the most compact size specification achieves a remarkable inference time of only 3.8 ms per prediction on an STM32 microcontroller. The proposed model is
$47 \times $
faster with
$16 \times $
fewer parameters compared to the ResNet architecture-based model. These results emphasize the significant potential of utilizing UWB impulse radar systems for practical people counting applications in real-world IoT scenarios, whether in static setups for indoor use or in a dynamic state mounted on an unmanned aerial vehicle.