New Efficient Depth Estimation and Real-Time Object 3D Recognition Models for Humanoid Robotic Environment Understanding

As demand for 3D visual capabilities in humanoid robots increases, the limitations of traditional methods are becoming evident. Point-cloud techniques, while effective for 3D reconstruction, are hindered by computational complexity and extensive data collection. Conversely, monocular depth estimation networks (mono-Net) are simpler but only suitable for specific VR applications, struggling to accurately measure real-world 3D coordinates. Both approaches can be slow for real-time use. To overcome this, we developed an efficient multi-object 3D detection system using coordinate mapping and object-scale aware median filtering. Our simulations show that this system outperforms existing methods, achieving over 90% accuracy and a 0.066-second inference time on an RTX-3060 GPU laptop, making it ideal for visual Edge-AI applications in logistic robots and cruise robots.