基于粒子滤波器的运动人手3D跟踪

3D hand tracking based on particle filtering

为改善基于粒子滤波算法的运动人手跟踪精度,结合高斯分布改进粒子滤波算法的采样过程.利用第k帧手势的帧图像信息和k-1帧手势的量测值恢复第k帧手势的量测值,按照高斯分布进行粒子采样,并根据手势特征点信息与手势的观测值衡量采样粒子的权重,调整粒子分布,然后根据运动过程中各个关节的预测值和真实值更新高斯分布的均值、方差,产生更接近真实状态的粒子.为验证算法的有效性,分别从跟踪精度和平均运行时间2个方面进行了大量对比实验.结果表明:与传统的粒子滤波算法相比,改进后的粒子滤波算法可以运用较少的粒子数达到较高的跟踪精度,节省了时间开销.引入高斯分布后,缩小了跟踪过程中手势真实解的搜索范围,避免了"维数灾难".

To improve the precision of 3D hand tracking based on the particle filtering,a novel algorithm using the Gaussian distribution to improve the sample process of particle filtering is put forward. The 3D hand gestures at time k are obtained by the frame information at time k and the measured values of the 3D hand gesture at time k 1. The particles are sampled according to the Gaussian distribution,and the weights of particles are obtained by establishing the mapping relation of features of 3D hand gestures and frame images. By adjusting the distribution of the particles and updating the mean value and the covariance in the Gaussian distribution according to the prediction and real state of each joint,the particles which are closer to the real states of hand gestures are obtained. To verify the effectiveness of the algorithm,a great number of experiments were carried out from the aspects of both tracking accuracy and average running time. The results show that the improved algorithm can take less running time to achieve better tracking precision and uses fewer particles compared with the conventional particle filtering method. By introducing the Gaussian distribution,the searching space is also reduced to avoid the'curse of dimensionality'.