Efficient rebar bundling: vision robotics innovations

ABSTRACT

The application of vision robots in rebar bundling operations has aroused great interest in the industry. The performance cost and complex application environment have become the key constraints to its development. Based on this, in this paper, we firstly design a rebar bundling intersection image acquisition system for rebar intersections in complex environments; secondly, we use GhostNet to lighten the backbone network and combine the Convolutional Block Attention Module (CBAM) and Bidirectional Feature Pyramid Network (BiFPN) with the neck to improve the adaptive ability to the complex environments; and finally, in order to make our network has better robustness, we try to introduce Distance Intersection Union (DIoU) in the network and verify the performance of the improved model through comparison and ablation experiments. The experimental results show that the mAP, accuracy and recall of the improved YOLOv5 model are 97.8%, 95.6% and 95.4%, which are 1.1%, 0.9% and 0.1% higher than that of YOLOv5s, respectively. Meanwhile, computational quantization and parameter quantization are reduced by 36% and 34%, respectively. This indicates that our proposed improved network has better intersection detection performance and also provides favorable conditions for deployment in robotic systems.

CO EDITOR-IN-CHIEF:

• Hsiau, Shu-San

ASSOCIATE EDITOR:

• Liu, De-Shin

KEYWORDS:

• Rebar bundling
• convolutional block attention module
• bi-directional feature pyramid network
• robotic systems

Nomenclature

FPS	=	It represents the number of frames of image or video processed in one second
$P_6^{in}$	=	It is the input of the $P_6$ layer
$P_5^{out}$	=	It is the output of the P_5 layer
$P_6^{td}$	=	It is the output of the $P_6$ intermediate layer
$P_7^{in}\mathit{\hspace{0.17em}}$	=	It is the input to the$\mathit{\hspace{0.17em}}{P}_7$ layer
mAP	=	It is a metric used to measure the performance of models doing object detection tasks
$w_1^{\prime },w_2^{\prime }$ and $w_3^{\prime }$	=	They are the weights of the inputs of the respective layers
$w_1$ and $w_2$	=	It is the respective learnable weight
$\mathit{\epsilon }$	=	It is taken as 0.0001 to avoid the denomi- nator being zero

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This study was supported by the National Natural Science Foundation of China under Grant No. [52265016] and related programs [2022D01C32 and 70222286]. The reinforcing steel equipment used was provided by the Xinjiang Academy of Building Science. The authors would like to express their sincerest gratitude to these organizations for making this study possible.