DMG6D: A Depth-based Multi-Flow Global Feature Fusion Network for 6D Pose Estimation
Zihang Wang, Qiang Zhang, Xueying Sun, Jianwei Zhu, Hao Wei · 2024
The accurate estimation of the 6-degree-of-freedom pose of a target object in the environment plays a pivotal role in robot perception, providing a foundation for interaction and manipulation between robots and the surrounding environment. Nonetheless, traditional vision sensors are prone to diminished reliability in visual perception due to environmental factors such as lighting conditions and occlusions. Depth sensors, such as Time-of-Flight (TOF) and structured light sensors, offer promising opportunities for reliable target pose estimation. However, accurately determining the pose solely based on a single depth image presents significant challenges due to the limited availability of rich appearance and texture information. To comprehensively address this challenge, we investigate the mechanism of feature extraction and representation using depth images, along with the utilization of normal angle and point cloud information derived from the depth images, to achieve robust estimation of the visual target poses. By exploiting the latent information within the depth images, including normal angles and point clouds, we have developed the DMG6D robust target pose estimation framework. Within the DMG6D framework, we first employ physical methods to infer the normal angle and spatial position of each pixel in the depth image. Subsequently, we introduce a three-branch feature extraction and a global feature fusion network to enable a comprehensive depiction of the target object. Finally, a robust pose estimation for the target object is obtained utilizing the least squares method. Experimental results emphatically demonstrate that the proposed DMG6D surpasses existing algorithms in terms of its ability to estimate 6D poses using depth images, effectively underscoring the efficacy of our designed depth image feature extraction strategy. Access to the code and video is available at https://github.com/wangzihanggg/DMG6D.