基于约束蒙特卡洛树搜索的大视场摄像机标定算法研究

Camera calibration is one of the essential problems in computer vision. Usually, a number of one dimensional or two dimensional small targets are combined into one large plane or stereo target to calibrate camera in large field of view. However, the problem of selecting the proper number and layout of the target according to the measuring object and environment still remains unsolved. In this project, firstly, the combination of small targets is taken as tree searching and Monte Carlo Tree Search (MCTS) algorithm is used to obtain approximate optimal solution. Secondly, image of large field of view is introduced as prior information to design objective function of MCTS, guide direction of reinforcement learning and enhance Monte Carlo simulation efficiency. Thirdly, nonlinear deep reinforcement learning model is established and model parameter learning and optimization algorithm is designed to balance exploration and utilization during search process. Finally, The leaf node of the highest calibration accuracy is taken as the final decision result to obtain the best scheme of number and layout of the small target. The research findings in this project have the virtue of fewer targets, high accuracy and strong adaptability in complex situations, which are expected to promote the development of image processing theory and calibration technology application and have important academic significance and practical value.

摄像机标定是计算机视觉领域研究的热点问题之一。目前，通常采用将多个小尺寸一维或二维靶标组合为大尺寸平面或立体靶标的方式来实现大视场环境下的摄像机标定。然而，如何根据测量对象和测量环境的特性选择合适的靶标数目和布局尚未得到有效解决。本项目首先将小尺寸靶标的组合归结为树形结构搜索的模式，利用蒙特卡洛树搜索算法（简称MCTS）得到近似最优解。其次，引入大视场图像的先验知识约束，设计MCTS目标函数，引导强化学习方向，提升蒙特卡洛模拟效率。再次，构建非线性深度强化学习模型，并设计模型参数学习和优化算法，平衡搜索过程中的探索和利用。最后，以标定精度最高的叶子节点作为最终决策结果，从而获取小尺寸靶标数目和布局的最佳方案。本项目的研究成果具有靶标数目少，对复杂环境适应性强，精度高等优点，将推动图像处理理论及应用标定技术的发展，具有重要的学术意义和实用价值。

摄像机标定是计算机视觉领域研究的热点问题之一。为了能够获得较高的标定精度，靶标需要尽可能覆盖整个视场。在大视场环境下，由于大靶标难以制作、运输和使用，因此过去通常采用将多个小尺寸一维或二维靶标组合为大尺寸平面或立体靶标的方式来实现大视场环境下的摄像机标定。然而，多个小尺寸靶标的共面性难以保证，这极大影响了传统方法的标定精度。.. 本项目首先基于约束MCTS思想，提出“虚拟大平面靶标”概念，并由此设计了一种大视场摄像机标定方法；其次，采用深度卷积神经网络，构建非线性深度强化学习模型，用以在扩展叶子节点前对各节点进行预评估，从而动态调整参数；最后，构建大视场摄像机标定系统，将所提方法与主流的标定方法进行对比测试。实验结果表明，所提方法不但标定精度在视场区域内具有更高的稳定性，而且具有其他方法不具备的靶标位置摆放灵活，操作简便等优点。.. 本项目已授权发明专利“一种基于四目立体视觉的直升机旋翼桨叶动态轨迹测量方法”和“一种基于标记点三维信息测量的飞机侧滑角动态监测方法”作为应用成果“大场景运动环境计算机视觉测量关键技术及应用”的重要组成部分，已应用于深圳市佳晨科技有限公司和深圳訾岽科技有限公司的相关产品中，提升了产品的性能，产生了良好的经济效益；这体现出本项目的研究成果具有广泛的应用前景。