无人机视觉下的输电塔线关键部件实时定位识别

UAV-Vision Based Real-Time Power Line Components Identification

摘要: 传统的高压输电线路巡检由专职检修人员负责，其劳动强度大、危险性高且效率较低。目前，基于无人机视觉的高压塔线系统巡检技术由于其成本低、安全性高、不受空间限制及航程远等优势在电力行业中受到了越来越多的关注和应用。无人机电力塔线系统巡检的首要任务是对输电塔线及其关键部件进行识别定位。目前，机器视觉中常用的目标检测算法以浅层结构模型为主(支持向量机、回归算法模型、传统BP神经元网络等)。而浅层机器学习模型受限于其本身的特征学习能力限制，在复杂背景中的识别效果较为有限。深度学习虽然具有较强的特征学习分类能力，但绝大多数深度学习中需要大量的卷积和微分计算，计算强度大耗时长。目前常用的基于深度学习的目标检测算法如RCNN等虽然在识别准确率上高于浅层算法，但识别过程需要的计算时间长，难以满足实时检测的要求。本文使用了最新的YOLO3深度学习目标检测算法模型对输电塔线系统中的关键部件进行识别取得了良好的效果，同时，由于YOLO3内只使用单个卷积神经元网络核心对视频图像中的目标检测对象进行识别，其效率更高速度更快。通过测试发现，基于YOLO3模型的输电塔线系统关键部件识别平均时间在0.36 ms，完全达到了实时检测的要求。

Abstract: Traditional power line inspections are mostly conducted manually with experienced workers. It is low efficient, labor intensive, and high-risk. Therefore, UAV vision based techniques have shown a growing interest in high-voltage power line inspection. The fundamental task for power line in-spection is the detection and classification of power components on the power transmission infra-structures. So far, classical machine learning based target recognition and detection algorithms, such as Support Vector Machine (SVM), Regression and other shallow structured machine learning models are difficult to achieve high accuracy. Deep learning models, convolutional neural network (CNN) for example, perform much better in target recognition and become the most preferred al-gorithm in many multi-target recognition scenarios. However, most CNN based target detection methods (RCNN, Faster RCNN, etc.) are computationally expensive. It is difficult to achieve real time detection in UAV-Video. YOLO3 (You Only Look Once V3) is recently developed CNN based target detection model, it has been proved as one of the top target recognition methods in terms of speed and accuracy. In this work, we deployed YOLO3 in power components recognition. The experiments show that the proposed method can be used to identify and locate power components within an average speed of 36 ms which fully achieves the speed requirement for real-time inspection.