MULTI-OBJECT GRASPING DETECTION BASED ON THE IMPROVED SHUFFLENET NETWORK

With the advent of the Industry 4.0 era, grasping technology has gradually become an essential skill of robot. Due to the existing problems, such as cross-domain adaptability, algorithm accuracy, speed, robustness to be improved, the lack of multi-object grasping datasets, and some networks have a poor performance in detecting small targets, this paper researches multi-object grasping detection based on the improved ShuffleNet network. In this paper, we independently make multi-target objects grasping dataset first. Then, We focus on the ShuffleNet and design the atrous spatial pyramid pooling (ASPP) + channel attention module and spatial attention module (CBAM) + CEASC + ShuffleNet (ACCSNet) based on the ShuffleNet model, and Adan is cited as the optimisation function when training the network. Finally, based on the constructed multi-target objects grasping dataset, the paper verify grasping experiments using the Kinova mico2 six-degree-of-freedom robotic arm in the complex multi- target scene. The experimental results show that the accuracy and speed of the ACCSNet are improved in the grasping process. Specifically, the experimental loss rate is only 1%, 4.8% lower than ShuffleNet, and the speed is 1 min faster than ShuffleNet each epoch.