液晶与显示, 2020, 35 (5): 464, 网络出版: 2020-05-30
基于神经网络与卡尔曼滤波的手部实时追踪方法
Hand real-time tracking method based on neural network and Kalman filter
摘要
针对传统的手部跟踪算法存在实时性差、识别精度低、易受环境影响等问题, 提出了一种基于神经网络与卡尔曼滤波的手部实时追踪方法。该方法首先通过神经网络对视频中出现的检测目标 进行定位, 接着用卡尔曼滤波对目标运动进行估计, 将估计的结果与下一帧图像中检测到的目标进行比对; 然后对检测到的目标进行跟踪, 将手部运动的轨迹实时显示。实验结果表明, 该方法能够 对多个手部目标实时进行追踪, 并在手部运动过程中出现交叉和形变的情况下还能保持跟踪, 其平均处理帧数为21.212 f/s, 追踪正确率为94.88%,基本满足手部跟踪的稳定可靠、高实时性、高鲁 棒性等要求。
Abstract
Aiming at the problems of poor hand-time tracking, low recognition accuracy and environmental impact, a hand tracking method based on neural network and Kalman filtering is proposed. The method firstly locates the detection target appearing in the video through the neural network, then estimates the target motion by Kalman filter, compares the estimated result with the detected target in the next frame image, and finally detects the target, tracking and displaying the trajectory of the hand movement in real time. Experiments show that the method can track multiple hand targets in real time and keep tracking when the cross and deformation occur during hand movement. The average processing frame number is 21.212 f/s, and the tracking accuracy rate is 94.88%. It basically meets the requirements of stable, reliable, high real-time and high robustness of hand tracking.
曾公任, 姚剑敏, 严群, 林志贤, 郭太良, 林畅. 基于神经网络与卡尔曼滤波的手部实时追踪方法[J]. 液晶与显示, 2020, 35(5): 464. ZENG Gong-ren, YAO Jian-min, YAN Qun, LIN Zhi-xian, GUO Tai-liang, LIN Chang. Hand real-time tracking method based on neural network and Kalman filter[J]. Chinese Journal of Liquid Crystals and Displays, 2020, 35(5): 464.