基于多传感器融合的三维目标跟踪

发布时间：2018-05-20 22:01

本文选题：三维目标跟踪 + 惯性测量传感器　；参考：《山东大学》2017年硕士论文

【摘要】：三维目标跟踪是计算机视觉领域的基础性问题,一直以来都是国内外学者们的研究热点。经典的基于模型的三维跟踪方法是将三维模型投影出的边与图像中的边进行对齐的配准过程,而利用图像处理技术可以得到三维模型投影到图像上的轮廓,然后在轮廓上均匀采样得到点,最后边的对齐实际上就是轮廓上的采样点与其法线方向上候选点的匹配。然而,这种低级的点匹配受光照变化、遮挡、运动模糊等因素的影响极易发生错误,导致跟踪的失败。为了解决低级匹配错误导致跟踪失败的问题,本文通过使用惯性测量传感器辅助视觉传感器来估计相机姿态,提出了一种能有效提高基于模型的三维目标跟踪鲁棒性的新方法,即基于多传感器融合的三维跟踪。主要贡献如下:1.通过在视觉传感器(相机)上绑定惯性测量传感器来估计相机的旋转。惯性测量传感器包含3轴陀螺仪、3轴加速度计和3轴地磁计,分别可以测量角速度、加速度和地磁强度。我们首先对从惯性测量传感器获得的原始数据使用卡尔曼滤波以降低测量误差。对角速度在时间上进行积分和递推,可以计算出转动角度,由此可以得到相机的旋转,然后为了提高相机旋转估计的精度我们提出基于补偿滤波的传感器融合算法。与此同时,通过对加速度的积分可以得到相机速度,这将为使用粒子滤波算法估计相机平移提供帮助。2.通过粒子滤波确定相机平移。把我们以已经得到的相机旋转和相机速度作为粒子滤波的输入,每个粒子表示潜在的一个相机的姿态,使用动态模型预测,观测模型修正。粒子只有相机平移三个自由度的状态空间,使得粒子滤波过程可以使用相对较多的粒子来保证跟踪效果,同时又能将计算代价控制在可接受范围之内。3.实现算法集成框架。本文将惯性测量传感器和粒子滤波集成到一个算法框架之下。通过本文提出的方法,可以实现稳定的跟踪。综上所述,本文就三维目标跟踪课题进行了研究,提出了一种通过使用多传感器融合进行跟踪的方法。通过实验,我们检验了所提出算法的可行性,并且通过定量分析和定性分析验证了算法具有很高的精确性和鲁棒性。
[Abstract]:3D target tracking is a fundamental problem in the field of computer vision, and has always been a hot research topic of scholars at home and abroad. The classical 3D tracking method based on model is to align the edge of the 3D model with the edge of the image, and the contour of the 3D model can be obtained by using the image processing technology. Then the point is uniformly sampled on the contour, and the alignment of the last edge is actually the matching between the sample point on the contour and the candidate point in its normal direction. However, the low-level point matching is affected by illumination, occlusion, motion blur and other factors, which leads to the failure of tracking. In order to solve the problem of tracking failure caused by low-level matching errors, this paper proposes a new method which can effectively improve the robustness of 3D target tracking based on model by using inertial measurement sensor aided visual sensor to estimate camera attitude. That is, three-dimensional tracking based on multi-sensor fusion. The main contributions are as follows: 1. The rotation of the camera is estimated by binding an inertial measurement sensor to the visual sensor (camera). The inertial measurement sensor consists of 3-axis gyroscopes, 3-axis accelerometers and 3-axis magnetometers, which can measure angular velocity, acceleration and geomagnetic intensity, respectively. We first use Kalman filter to reduce the measurement error of the raw data obtained from the inertial measurement sensor. In order to improve the accuracy of camera rotation estimation, a sensor fusion algorithm based on compensatory filtering is proposed in order to improve the accuracy of camera rotation estimation by integrating and recursive angular velocity in time, the rotation angle can be calculated and the rotation angle can be obtained. At the same time, the camera speed can be obtained by integrating the acceleration, which will help to estimate the translation of the camera using the particle filter algorithm. The camera translation is determined by particle filter. We use the camera rotation and camera speed as the input of particle filter. Each particle represents the attitude of a potential camera. The dynamic model is used to predict and the observation model is modified. The state space of three degrees of freedom is only shifted by the camera, which makes the particle filter process use relatively more particles to ensure the tracking effect, and at the same time, the computational cost can be controlled within the acceptable range of .3. The algorithm integration framework is implemented. In this paper, the inertial measurement sensor and particle filter are integrated into an algorithm framework. The method proposed in this paper can realize stable tracking. To sum up, this paper studies the subject of 3D target tracking, and proposes a method of tracking by using multi-sensor fusion. Through experiments, the feasibility of the proposed algorithm is verified, and the accuracy and robustness of the algorithm are verified by quantitative analysis and qualitative analysis.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP212

【相似文献】