基于改进SOLOv2的复杂场景下智能机器人巡检识别算法  

Inspection and identification algorithm based on improved SOLOv2 of intelligent robot in complex environment

作　　者：吴忧 袁雪[1] WU You;YUAN Xue(School of Electronics and Information Engineering,Beijing Jiaotong University,Beijing 100044,China)

机构地区：[1]北京交通大学电子信息工程学院,北京100044

出　　处：《北京交通大学学报》2022年第5期95-106,共12页JOURNAL OF BEIJING JIAOTONG UNIVERSITY

基　　金：国家自然科学基金(61871024)。

摘　　要：智能巡检技术逐步替代人工巡检有效提高检测效率,降低人力成本,成为智能巡检系统中的重要组成部分.目前现有的智能巡检技术在复杂场景下可识别的设备部件种类较少,检测精度和速度较低,对于小部件(例如按钮等)的识别效果较差.为了达到较好的检测效果和一定程度的广泛适用性,提出了一种基于改进SOLOv2的智能机器人巡检识别算法.本方法能较好地应对复杂场景下多种类设备部件的精准定位及模态提取问题,解决了现有智能巡检技术可识别部件种类少的问题.针对实例分割算法SOLOv2对于小目标识别精度低的问题,通过增加特征金字塔网络中大尺寸层级特征图的输出,增加小目标物体的正样本数量,提高小目标识别精度.实验结果表明,本文提出的方法相较于目前的巡检识别算法,具有更好的识别精度,复杂场景下的鲁棒性更高;相比较原有的智能巡检系统,可识别种类提高12类;相较于原始SOLOv2算法,小目标物体的精度提升10%左右,整体的识别精度也提升1.7%.With the rapid development of artificial intelligence,intelligent inspection technology gradually replaces manual inspection,effectively improves inspection efficiency,reduces labor cost and becomes an important part of intelligent inspection system.At present,the existing intelligent inspection technology can identify fewer types of equipment components in complex scenarios and the detection accuracy and speed are relatively low.Small parts(such as buttons,etc.)have poor recognition performance.In order to achieve better detection effect and a certain degree of wide applicability,an intelligent robot inspection and identification algorithm based on improved SOLOv2 is proposed.This method can better deal with the problem of accurate positioning and modal extraction of various types of equipment components in complex scenes and solves the problem that the existing intelligent inspection technology can identify few types of components.Aiming at the problem that the instance segmen‐tation algorithm SOLOv2 has low recognition accuracy for small objects,by adding the output of large-size hierarchical feature maps in the feature pyramid network,the number of positive samples of small target objects is increased and the recognition accuracy of small objects is improved.The experi‐mental results show that the method proposed in this paper has better recognition accuracy and higher robustness in complex scenes than the current inspection and identification algorithm;compared with the original intelligent inspection system,the number of identifiable components increases 12 catego‐ries;compared with the original SOLOv2 algorithm,the accuracy of small target objects is improved by about 10%and the overall recognition accuracy is also improved by 1.7%.

关 键 词：智能巡检 实例分割 SOLOv2 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]