Multi-object recognition and grasping detection based on the anchor-free network.

Mingming Meng,Weihai Chen,Xiantao Sun,Wenjie Chen

DOI: https://doi.org/10.1109/CISP-BMEI60920.2023.10373320

2023-01-01

Abstract:In multi-object scenarios, problems such as low accuracy of detection algorithms and complex network structures arise when a robotic arm is intended for grabbing specific objects. To address these issues, we propose a multi-task network model based on the anchor-free mechanism to detect the category and grasping pose of objects. Firstly, using a residual network, a shared main trunk feature extraction module is designed for shallow feature extraction used in object detection and grabbing detection branches. Secondly, a grasping strategy based on key-point is introduced, which can obtain an accurate estimate of grabbing pose through the key-point heatmap. Lastly, in the grabbing prediction module, a hybrid attention mechanism is introduced to improve the grabbing prediction accuracy. Experimental results show that the detection accuracy of this method on the multi-target grasping data set VMRD is 94.2%. In the real multi-target scene, the target detection and grasping detection accuracy are 97.9% and 92.8% respectively.