基于单目序列影像的运动目标位置姿态实时监控方法研究

Real-time position and posture monitoring of moving objects has important significance in the field of military and civil application. For example, real-time position and posture monitoring of fighter planes is important for drawing up operation plan and quick response; real-time position and posture monitoring of planes or vehicles in airport or parking lot is important for scheduling management and route plan. Though the method of determination of position and posture based on stereo vision is strict in theory and is mature in implementation, it is strict in the demand of distribution of camera which many real applications can’t meet. This project deeply analyzes the motion continuity of space-time and the constraint provided by known and fixed environment which intrinsically exist in motion, proposes moving objects position and posture determination method based on monocular image sequences to conduct theory correction and method research pointing to the problem of real-time position and posture monitoring of moving objects in known environment. For this purpose, this project will learn from the thought of generalized point photogrammetry and Kalman filter to propose a new method to measure position and posture of moving objects on real-time. The method is modeled by using the motion continuity of space-time and the constraint of the known and fixed environment, in which the real-time feature is guaranteed by the thought of motion parameter updating and algorithm parallelization. Such research can’t only be widely applied to various practices, but also can extend and improve current theory of real-time monitoring of moving objects.

运动目标位置姿态的实时监控在军事和民用领域都具有重要意义，比如实时监控舰艇甲板上各战机的位置姿态对于及时制定作战计划和做出快速应对反应至关重要，对民用机场或停车场内的飞机或车辆的位置姿态实时监控对于调度管理和路径规划等具有重要作用。基于双目立体视觉的定位定姿方法理论严密，方法成熟，但是该方法对摄像机布设条件要求苛刻，很多实际应用无法满足。本项目针对运动分析的特点，深入分析运动本身具有的时空连续性以及已知环境条件所提供的约束，提出基于单目序列影像的运动目标定位定姿方法，对已知环境运动目标位置姿态的实时监控问题进行理论完善和方法研究。本项目汲取广义点摄影测量以及卡尔曼滤波的状态更新思想，基于运动时空连续性以及固定环境约束条件进行数学建模，利用状态更新和算法并行等手段，实现基于单目影像的针对已知环境运动目标位置姿态实时测量的方法。该方法操作简便、具有较高的稳定性可靠性。

单目序列影像是计算机视觉领域的重要数据源和研究对象，被大量用于目标识别、人-车流量统计、动作捕捉、危险预警、变化监测、三维重建等应用领域。本项目研究了运动目标位置姿态测量的理论和方法，在对相机进行内参数标定、外参数标定以及已知运动目标的三维形状的条件下，利用单目影像对运动目标进行较为精确的定位定姿。在具有复杂背景的影像中，利用全局和局部信息进行运动目标主动轮廓提取，利用高斯混合模型阈值法提取目标的完整轮廓，建立了影像中的目标轮廓与三维模型轮廓的吻合度评价指标，通过NEWUOA方法实现三维模型位置姿态参数的非线性迭代优化，利用扩展的卡尔曼滤波方法对连续测量的位姿数据进行分析预测，最终得到运动目标位姿姿态的估计值。本项目研究了不同场景下布设多个相机对场景实现全覆盖的设计方法，利用OpenGL技术建立了虚拟现实仿真环境，并且利用多个真实相机和模型建立了1:50比例缩小真实实验环境，多方面验证了方法的有效性。在效率方面，研究了基于分布式计算和GPU的实时处理方法，达到了优于10帧/秒的监测速度。基于本项目建立的理论和方法，还创新性应用到了无人机机载单目相机拍摄序列影像，对电信天线盒子、风力发电机叶片等固定目标的位置和姿态测量，得到了良好的效果。