Light-weight Monocular Depth Estimation Via Cross Attention Fusion of Sparse LiDAR

This article proposes a light-weight monocular depth estimation model applicable to mobile robots. Unlike autonomous vehicles, mobile robots face constraints in sensor and computing resources owing to considerations of a power efficient and light-weight design. Considering these constraints, we propose a model that estimates depth images from small camera images with minimal parameters and computational overhead. Additionally, to address the performance degradation that occurs during the model’s light-weighting process, we efficiently integrate sparse LiDAR point cloud through cross-attention mechanisms. This enables mobile robots to effectively acquire depth information about their surroundings.