Determination of relative three-dimensional (3D) position, orientation, and relative motion between two reference frames is an important problem in robotic guidance, manipulation, and assembly as well as in other fiel...Determination of relative three-dimensional (3D) position, orientation, and relative motion between two reference frames is an important problem in robotic guidance, manipulation, and assembly as well as in other fields such as photogrammetry. A solution to pose and motion estimation problem that uses two-dimensional (2D) intensity images from a single camera is de- sirable for real-time applications. The difficulty in performing this measurement is that the process of projecting 3D object features to 2D images is a nonlinear transformation. In this paper, the 3D transformation is modeled as a nonlinear stochastic system with the state estimation providing six degrees-of-freedom motion and position values, using line features in image plane as measuring inputs and dual quaternion to represent both rotation and translation in a unified notation. A filtering method called the Gaussian particle filter (GPF) based on the particle filtering concept is presented for 3D pose and motion estimation of a moving target from monocular image sequences. The method has been implemented with simulated data, and simulation results are provided along with comparisons to the extended Kalman filter (EKF) and the unscented Kalman filter (UKF) to show the relative advantages of the GPF. Simulation results showed that GPF is a superior alternative to EKF and UKF.展开更多
为了方便乒乓球运动训练和相关体育赛事评分,提出一种乒乓球姿态动作评分方法。通过深度学习神经网络提取骨骼关节点数据,得到一组动作序列的空间角度变化值;通过改进的动态时间规整DTW(dynamic time warping)算法把两组动作序列的关节...为了方便乒乓球运动训练和相关体育赛事评分,提出一种乒乓球姿态动作评分方法。通过深度学习神经网络提取骨骼关节点数据,得到一组动作序列的空间角度变化值;通过改进的动态时间规整DTW(dynamic time warping)算法把两组动作序列的关节点数据进行匹配,得到序列之间的距离;根据Mean Shift算法聚类数决定各个骨骼关节点权重系数,计算得到两个动作序列距离的大小评分值。实验表明,该方法能很好地实现人体动作评分,提高了关键关节点权重分值的影响,提高了乒乓动作序列评分的准确度,更加符合专家主观评分标准。展开更多
为提高群体活动场景下细粒度人体姿态估计的准确率,优化网路中人体识别及姿态估计算法,在现有研究的基础上,提出一种结合多尺度预测以及改进并行注意力模块的多目标人体姿态估计算法。在充分利用不同尺度特征信息的基础上,实现高质量的...为提高群体活动场景下细粒度人体姿态估计的准确率,优化网路中人体识别及姿态估计算法,在现有研究的基础上,提出一种结合多尺度预测以及改进并行注意力模块的多目标人体姿态估计算法。在充分利用不同尺度特征信息的基础上,实现高质量的人体姿态估计;针对运动场景下多目标人体姿态数据集较少,提出一种数据集CUPB Sport Dataset。实验结果表明,该算法在公开基准数据集和自制数据集上分别达到了81.4 mAP和79.7 mAP,验证了该算法在运动场景下针对多目标的高效性。展开更多
基金Project (No. 2006J0017) supported by the Natural Science Foundation of Fujian Province, China
文摘Determination of relative three-dimensional (3D) position, orientation, and relative motion between two reference frames is an important problem in robotic guidance, manipulation, and assembly as well as in other fields such as photogrammetry. A solution to pose and motion estimation problem that uses two-dimensional (2D) intensity images from a single camera is de- sirable for real-time applications. The difficulty in performing this measurement is that the process of projecting 3D object features to 2D images is a nonlinear transformation. In this paper, the 3D transformation is modeled as a nonlinear stochastic system with the state estimation providing six degrees-of-freedom motion and position values, using line features in image plane as measuring inputs and dual quaternion to represent both rotation and translation in a unified notation. A filtering method called the Gaussian particle filter (GPF) based on the particle filtering concept is presented for 3D pose and motion estimation of a moving target from monocular image sequences. The method has been implemented with simulated data, and simulation results are provided along with comparisons to the extended Kalman filter (EKF) and the unscented Kalman filter (UKF) to show the relative advantages of the GPF. Simulation results showed that GPF is a superior alternative to EKF and UKF.
文摘为了方便乒乓球运动训练和相关体育赛事评分,提出一种乒乓球姿态动作评分方法。通过深度学习神经网络提取骨骼关节点数据,得到一组动作序列的空间角度变化值;通过改进的动态时间规整DTW(dynamic time warping)算法把两组动作序列的关节点数据进行匹配,得到序列之间的距离;根据Mean Shift算法聚类数决定各个骨骼关节点权重系数,计算得到两个动作序列距离的大小评分值。实验表明,该方法能很好地实现人体动作评分,提高了关键关节点权重分值的影响,提高了乒乓动作序列评分的准确度,更加符合专家主观评分标准。
文摘为提高群体活动场景下细粒度人体姿态估计的准确率,优化网路中人体识别及姿态估计算法,在现有研究的基础上,提出一种结合多尺度预测以及改进并行注意力模块的多目标人体姿态估计算法。在充分利用不同尺度特征信息的基础上,实现高质量的人体姿态估计;针对运动场景下多目标人体姿态数据集较少,提出一种数据集CUPB Sport Dataset。实验结果表明,该算法在公开基准数据集和自制数据集上分别达到了81.4 mAP和79.7 mAP,验证了该算法在运动场景下针对多目标的高效性。
