A Novel RGB-D Cross-Background Robot Grasp Detection Dataset and Background-Adaptive Grasping Network

The existing industrial and household robots suffer from insufficient adaptability under increasingly complex scenes and challenges in operating objects. This prevents the manufacturing industry from flexibly responding to environmental changes and has become a pressing issue in current research. To address the aforementioned challenges, we propose a novel RGB-D cross-background robot grasp detection dataset (CBRGD). The dataset aims to enhance the performance of intelligent robots when confronted with diverse and dynamic scenarios. The CBRGD dataset consists of seven different common backgrounds and 58 common objects, covering various grasping scenarios. Alongside the dataset, we also propose a background-adaptive robot multiobject grasp detection network named BA-Grasp. Using the bimodal foreground activation strategy (BFAS), we can suppress background noise while highlighting the target object in the foreground. The multiscale foreground region features enhancement module (MFRFEM), on the other hand, allows the network to adapt to objects of different sizes while highlighting important areas in the foreground. Extensive experiments indicate that our robotic grasping method demonstrates significant advantages when facing various scenarios and complex, variable objects. The proposed method achieved the accuracies of 97.8% and 96.7% on the publicly available Cornell dataset and 94.6% on the Jacquard dataset, which are comparable to the SOTA method. Moreover, on the CBRGD dataset, the proposed method demonstrated an average accuracy improvement of 5%–10% over the SOTA method. Robot grasping experiments validate that our method can maintain high grasping accuracy in real-world applications. Experimental videos and the CBRGD dataset can be found at the following address:
https://github.com/meiguiz/BA-Grasp
.