Efficient Mobile Robot Navigation Using Quantized Monocular Surface Normal Estimation

Surface normal provides rich geometric information for robot navigation. Whether avoiding obstacles, traversing slopes, improving grasping, or enhancing SLAM, incorporating surface normal leads to more intelligent and adaptive robotic systems. LiDAR-based surface normal estimation is bulky and expensive, whereas camera-based CNN approaches offer a more practical and cost-effective solution. This paper presents an experimental evaluation of dynamically quantized weights in a monocular depth estimation model for identifying traversable and non-traversable areas using surface normal analysis in robotic navigation. By optimizing monocular depth-based surface normal estimation with application of median filtering step to remove salt-and-pepper noise, the qualitatively assessed quantized version of the pre-trained MonoDepth-PyTorch model generates surface normals suitable for indoor ROS-based mobile robot navigation.