Improving 3D Human Body Reconstruction Using Millimeter-Wave Radar-Based Kansformer

As the unique advantages of millimeter wave (mmWave) radar technology in handling complex environments become increasingly apparent, research into constructing 3D human models using mmWave radar has received extensive attention. Capable of operating effectively in adverse conditions such as smoke, snow, and poor lighting, mmWave radar holds significant potential for applications in emergency response and security monitoring. This paper investigates the possibility of reconstructing accurate 3D skeletons or mesh models by analyzing noisy and sparse radar signals. Moreover, addressing the issues of large parameter size and long training periods inherent in traditional Transformer models, this study employs an optimized Kolmogorov-Arnold Network(KAN) [1] structure to enhance the Transformer model, which called Kansformer. This structural optimization significantly reduces the model’s parameter count and improves its learning efficiency and generalizability across different scenes. The results demonstrate that the optimized model exhibits higher robustness and accuracy in mmWave human body reconstruction tasks, especially in visually obstructed environments.