Deep Learning-based Mobile Robot Target Object Localization and Pose Estimation Research

Two key technologies in robotic object grasping are target localization and pose estimation (PE), respectively, the addition of a vision system can dramatically enhance flexibility accuracy grasping. The study optimizes classical convolutional structure detection network considering limited computing power memory resources embedded platform, replaces original anchor frame mechanism using an adaptive combination with fused depth map. For evaluating target’s pose, smooth plane its surface is identified semantic segmentation network, information obtained by solving normal vector plane, so that arm absorb along direction to achieve maintain average 85.75% even when number frames increased, which proves anti-interference ability over fitting problem. algorithm 98.8%; PE 74.32%; operation speed could be 25 frames/s. It satisfy requirements real-time physical In view study, experiments were carried on. Then success rate was above 75%, effectively verified practicability.