A Real-time Explainable-by-design Super-Resolution Model for LiDAR SLAM

Addressing the prohibitive cost of high-resolution LiDAR sensors, this study introduces a novel LiDAR Super-Resolution method. In a more detailed manner, we will emphasize on three key factors for deploying Super-Resolution algorithms in real-world scenarios: enhanced accuracy, reduced computational complexity, and real-time performance. To this end, a novel optimization problem is proposed to solve the LiDAR Super-Resolution problem. By combining learnable regularizers that utilize the low-rank properties of the LiDAR data and handcrafted regularizers that capture the ring-like structure, an explainable and computationally efficient deep learning architecture is derived using the Deep Unrolling strategy. Extensive numerical results on a real-world autonomous driving dataset demonstrate our model's superiority over existing state-of-the-art methods, especially in practical SLAM applications. Notably, our solution operates at 50-100 fps, significantly improving realtime performance, a key advantage over current methods that do not meet this critical requirement.