This paper aims to address the problem of complex accurate segmentation of target objects due to chaotic stacking of multiple objects of the same kind in the vision-based robotic arm order review scenarios. This paper proposes an improved fully convolutional one-stage object detection (FCOS) rotation detection network for order review.
An improved FCOS rotation detection network for the vision-based robotic arm order review scenario is proposed. This method solves the problem of difficulty in accurately segmenting multiple objects when overlapped and stacked by introducing a sliding window multi-head self-attention (SW-MSA) and angle parameters. Secondly, the AdamW optimization algorithm is introduced to obtain an adaptive learning rate and improve the training efficiency of the network. Thirdly, the Rotation IoU Loss is introduced as the loss function for bounding box regression to enhance the model’s precise positioning of the rotation target position.
In the same conditions, the proposed improved FCOS rotation detection network was trained for 12 epochs, which took 121 s less than the original FCOS detection network, a speed increase of 22.9%. With two images loaded at a time, the memory used was reduced by 254 MB. The detection speed also increased from 2.5 to 3.4 images per second. Finally, the comparative and ablation experiments on the DOTA-v1.0 data set and the self-made data set further verified through experiments that the improved FCOS rotation detection network is practical and effective in terms of detection accuracy and operational efficiency in real-world environments.
An improved FCOS rotation detection network for the robotic arm in the order review scenario is proposed. This method effectively solves the problem of inefficient segmentation of the target object caused by the chaotic stacking of multiple similar objects.
