【作者】 刘志飞；

【导师】 刘健波；

【作者基本信息】 四川大学 ， 工程硕士（专业学位）， 2024， 硕士

【摘要】 相机在计算机视觉应用例如自动驾驶、三维重建和虚拟现实等领域扮演着重要的角色,同时,作为计算机视觉领域的基础组件,相机模型及其标定也是其重要组成部分,它起到了对图像和真实世界的映射作用,是后续许多下游任务必不可少的预处理步骤。相机模型标定误差会传导到下游任务并影响其性能,因此,真实地建模相机模型并对其进行准确标定具有重大研究意义和应用价值。当前大部分参数化相机模型很难完全拟合真实镜头成像过程中的畸变,而基于Raxel的相机模型虽然表达能力很强,但其参数量很大,标定困难,实用价值不高,并且几乎所有基于Raxel的相机模型都没有考虑实际成像系统几何。针对当前相机模型及其标定中存在的问题与局限性,本文主要做了以下相关工作:(1)离线相机标定中传统棋盘格角点检测算法只能在所有角点均检测到时才能恢复角点匹配信息,并且无法检测遮挡或部分标定板图像,同时其角点处梯度信息有限,使得其角点检测精度有限。此外,大多数方法使用固定大小窗口进行角点细化,这种方法无法适应畸变较大或者深度范围较广的标定板图像。因此,本文提出一种基于自适应窗口大小的标定板角点检测算法,使用对称星形棋盘格和AprilTag组成的标定板以增加角点梯度信息和标定板检测率,并使用基于像素强度匹配和对称的方法来细化角点。实验结果表明,该方法可以达到更高的角点检测率和图像检测范围,同时有着很高的角点检测精度,这对后续相机标定准确性和稳定性十分重要。(2)提出一种基于近似焦散约束的Raxel相机模型及其标定方法。该相机模型基于Raxel模型,但使用降采样的raxel来降低模型参数,同时使用迭代优化算法估计投影亚像素位置,对于反向投影则使用三阶B样条插值来估计给定亚像素位置处的raxel。针对实际成像系统焦散求解过程复杂、计算量大且无法利用稀疏角点匹配信息进行求解的问题,本文提出近似焦散的概念,可以容易在相机模型参数优化过程中对相机模型的raxel应用近似焦散约束,使得可以利用稀疏角点匹配信息来标定相机模型,同时使优化后的raxel有更好的物理可解释性。在多个数据集上的对比实验结果从误差、偏差、参数量和可解释性等方面表明了本文方法的优越性。更多还原

【Abstract】 Cameras play an important role in computer vision applications such as autonomous driv-ing,3D reconstruction and virtual reality.As fundamental components in the field of computer vision,camera models and camera calibration are crucial.They serve the purpose of map-ping images to the real world and are indispensable preprocessing steps for many downstream tasks.Errors in camera calibration propagate to downstream tasks and affect their performance.Therefore,accurately modeling camera models and performing precise camera calibration hold significant research and practical value.In the one hand,most parametric camera models cur-rently available struggle to fully capture the distortion present in real lens imaging processes.On the other hand,Raxel-based camera models have high expressive power but come with a large number of parameters,making calibration difficult and limiting their practical utility.Addi-tionally,almost all Raxel-based camera models do not consider the geometry of actual imaging systems.In response to the issues and limitations of current camera models and camera calibration,this thesis focuses on the following related works:(1)Traditional chessboard corner detection algorithms used in offline camera calibration can only recover corner matching information when all corners are detected.They are unable to handle occlusion or partially appeared board images.Furthermore,the limited gradient in-formation at corner locations results in limited accuracy of corner detection.Additionally,most methods use a fixed-size window for corner refinement,which is not suitable for board images with significant distortion or a wide depth range.Therefore,this paper proposes a corner de-tection algorithm for calibration board based on adaptive window size.It utilizes symmetric star-shaped chessboard and AprilTags to enhance corner gradient information and board detec-tion rate.The proposed method employs pixel intensity matching and symmetric approaches for corner refinement.Experimental results demonstrate that this method achieves higher corner detection rate and image detection areas,while maintaining high accuracy.This is crucial for the accuracy and stability of subsequent camera calibration.(2)This paper proposes a camera model and calibration method based on the approximated caustic constraint.The camera model is based on the Raxel model but employs downsampled raxels to reduce the model parameters.It utilizes an iterative optimization algorithm to estimate sub-pixel positions for projection,and employs third-order B-spline interpolation for estimat-ing raxels at given sub-pixel positions during back projection.To address the challenges of complex solving processes,high computational load and the inability to utilize sparse corner matching information on the computing of the caustic,the concept of approximated caustic is introduced.This concept allows for the application of approximated caustic constraints to raxels in the camera model parameter optimization process,enabling the utilization of sparse corner matching information for camera calibration.It also improves the physical interpretability of the optimized raxels.Comparative experimental results on multiple datasets demonstrate the superiority of the proposed method in terms of error,bias,parameter number,and interpretabil-ity.更多还原