Palletizing Robot Positioning Bolt Detection Based on Improved YOLO-V3

To improve the detection accuracy and speed of palletizing robot positioning bolts in complex scenes, we proposed a positioning bolt (PB) detection method based on improved YOLO-V3. First, due to the actual detection requirement, we constructed the PB data set by using a series of data enhancement operations such as horizontal flip, +/- 30degree rotation, and random luminance enhancement or decrease. Then, an improved anchor box mechanism based on the k-means++ algorithm was designed to obtain a more accurate anchor box for the PB data. According to the feature of the PB data in the palletizing robot, such as the existence of dust and dirt on the surface, the feature extraction network was further enhanced by adding a Densenet-4 module. In this way, the low-level semantics and high-level abstract features can be extracted effectively to improve detection performance. Finally, a new bounding box regression loss function was elaborated to accelerate the neural network training. The experimental results demonstrated the effectiveness of the proposed improvement mechanisms. The comparable results also show that our method is superior to the original YOLO-V3, SSD, and Faster R-CNN for PB data, and has a detection AP of 86.7%, a recall rate of 97%, and a detection speed of 25.47 FPS, which can achieve high-efficiency and high-precision detection in complex industrial scenarios.

Automated summary

This paper proposes an improved method for detecting positioning bolts (PB) by enhancing the dataset, designing an improved anchor box mechanism, and enhancing the feature extraction network. The method improves the detection accuracy and speed of PB data in palletizing robots.