一种基于改进YOLOv7的相机标定特征点检测方法  

A method for camera calibration feature point detection based on an improved YOLOv7

作　　者：陈松[1,2] 闫国闯 马方远 王西泉 田晓耕 CHEN Song;YAN Guochuang;MA Fangyuan;WANG Xiquan;TIAN Xiaogeng(State Key Laboratory for Strength and Vibration of Mechanical Structures,Xi’an Jiaotong University,Xi’an 710049,China;Norinco Group Testing and Research Institute,Xi’an 710116,China)

机构地区：[1]西安交通大学复杂服役环境重大装备结构强度与寿命全国重点实验室,西安710049 [2]中国兵器工业试验测试研究院,西安710116

出　　处：《兵器装备工程学报》2024年第8期151-160,共10页Journal of Ordnance Equipment Engineering

基　　金：国家级科研项目。

摘　　要：在基于视觉方法的军事目标检测等技术中,相机的精确标定是进行目标高精度测量的前提,同时也是开展后续图像处理、目标跟踪、三维重建的基础。相机标定的关键在于准确的检测图像中的标定特征点。以当前使用范围较广的棋盘格标定法为对象,针对受干扰(模糊、重噪声、极端姿态和大镜头失真)的标定图像难以进行特征点提取的问题,提出一种融合改进YOLOv7-tiny深度学习网络和Harris角点检测的相机标定特征点检测算法。针对原始网络在相机标定特征区域检测中的各种问题,引入Gather-and-Distribute信息聚合分发机制替换YOLOv7-tiny的加强特征提取网络(FPN)部分,提高不同层之间特征融合的能力;在主干特征提取部分后加入Biformer注意力机制,提高对小尺寸特征点候选区域的捕捉能力;在Head部分使用改进Efficient Decoupled Head解耦头,在提高精度的同时维持了较低的计算开销。测试结果表明,改进后的YOLOv7-tiny网络对特征点候选区域检测的准确率有显著的提高,达到95.3%,证明了改进后网络的有效性和可行性。In military target detection technologies based on visual methods,precise camera calibration is essential for accurate target measurements and serves as the foundation for subsequent image processing,target tracking,and 3D reconstruction.The crux of camera calibration lies in the accurate detection of calibration feature points in images.Focusing on the widely-used chessboard calibration method,this paper addresses the difficulty in feature point extraction from calibration images affected by disturbances such as blur,heavy noise,extreme poses,and significant lens distortion.We propose a camera calibration feature point detection algorithm that integrates an enhanced YOLOv7-tiny deep learning network with Harris corner detection.To address various issues in the original network’s detection of camera calibration feature regions,we introduce a Gather-and-Distribute information aggregation and distribution mechanism to replace the Feature Pyramid Network(FPN)in YOLOv7-tiny,enhancing the capability of feature fusion across different layers.Additionally,a Biformer attention mechanism is added after the main feature extraction segment to enhance the detection of small-sized feature point candidate regions.In the Head section,an improved Efficient Decoupled Head is used to increase accuracy while maintaining low computational overhead.Test results demonstrate that the improved YOLOv7-tiny network significantly enhances the accuracy of feature point candidate region detection,achieving 95.3%,thereby proving the effectiveness and feasibility of the enhanced network.

关 键 词：相机标定 深度学习 YOLOv7-tiny 信息聚合分发机制 注意力机制 HARRIS算法 

分 类 号：TP24[自动化与计算机技术—检测技术与自动化装置] TP391.4[自动化与计算机技术—控制科学与工程]