冷库中基于YOLOv5的识别同伴机器人方法

A Method for Identifying Companion Robots in Cold Storage Based on YOLOv5

摘要: 针对现有目标识别算法因冷库低光照、货物遮挡等因素造成识别丢失的问题，提出一种基于改进的YOLOv5目标模型。首先，使用AutomatedMSRCR算法进行对采集到的光照不足图片图像预处理，提高图像质量。其次，采用kmeans++算法结合双线性插值法，重新设计锚框尺寸，为被遮挡区域填充合理的像素值。最后，为解决同伴机器人运动时尺度变化导致的识别差，融合CoordConv卷积提高网络对空间信息和尺寸的敏感度，提升模型的识别准确率。经实验表明，相比于现有YOLOv5模型，本文改进YOLOv5模型的mAP0.5、F1、精确率、召回率、分别提升4.5%、3.1%、3.5%、2.7%，满足实际应用场景中协作机器人对同伴准确识别的要求。

Abstract: A modified YOLOv5 object model is proposed to address the issue of recognition loss caused by fac-tors such as low lighting in cold storage and obstruction of goods in existing object recognition algo-rithms. Firstly, use the AutomatedMSRCR algorithm to preprocess the collected images with insuffi-cient lighting to improve image quality. Secondly, using the kmeans++ algorithm combined with bi-linear interpolation, the anchor box size is redesigned to fill the occluded area with reasonable pixel values. Finally, to address the recognition errors caused by scale changes during the movement of companion robots, CoordConv convolution is integrated to improve the network’s sensitivity to spa-tial information and size, and to improve the recognition accuracy of the model. Through experi-ments, it has been shown that compared to the existing YOLOv5 model, this improved YOLOv5 model improves mAP0.5, F1, accuracy, and recall by 4.5%, 3.1%, 3.5%, and 2.7%, respectively, meeting the requirements of collaborative robots for accurate peer recognition in practical applica-tion scenarios.