A millimeter-wave radar data optimization algorithm based on cross-modal learning

With the rapid development of sensors, there is an increasing demand for their perception performance in complex environments. Currently, the main types of environmental perception sensors are visual sensors, LiDAR sensors, and millimeter-wave radar sensors. LiDAR can provide high-precision point cloud data under normal conditions, but its performance is limited in complex environments, such as smoke. Millimeter-wave radar has higher robustness in harsh weather and low visibility conditions, but its point cloud data quality and precision are significantly lower than LiDAR due to limitations in sensor performance. To address this issue, this paper proposes a cross-modal learning-based millimeter-wave radar data optimization algorithm. The radar is trained using a diffusion model. Building upon the original model, we introduce an innovative hierarchical attention mechanism. By combining local self-attention and global cross-attention, the local self-attention captures detailed features, while global cross-attention enhances contextual fusion. This addresses the issue of detail loss and context integration in traditional attention mechanisms under high-resolution conditions. Compared to the original attention mechanism, the proposed approach reduces computational complexity and hardware memory usage while maintaining accuracy, and it improves processing speed. Additionally, based on the model's parameter characteristics, the latest Lion learning rate optimizer is selected to replace the original RAdam optimizer, which accelerates training speed and reduces the number of steps required for convergence, thereby improving training efficiency. Experimental results demonstrate that the proposed method can increase the number of millimeter-wave radar point clouds by 10%.