基于实时手势识别与跟踪的人机交互实现

为了实现对手势目标的自动识别和连续跟踪,提出了一种手势识别与跟踪算法。首先,通过离线训练手势目标检测器来实现手势目标的自动识别。接着,通过改进的Shi-Tomisi算法,在目标区域提取可靠稳定的特征点。然后,通过KLT跟踪器对特征点进行跟踪。当特征点跟踪成功时,通过求解仿射变换矩阵确定手势目标的新位置;当目标出现遮挡和大尺度旋转时,特征点丢失,此时在KLT跟踪器中加入卡尔曼滤波器来预测手势目标的位置,实现对手势目标的连续跟踪。同时对手势目标可能存在的区域进行估计,缩小检测器的检测范围,提高检测速度。最后,将算法应用于人机交互系统中,实现了机器人的远程控制。实验结果显示,算法在简单背景下的跟踪正确率为99.54%,复杂背景下的跟踪正确率为98.24%。实验结果表明,算法能够快速准确地对手势目标进行检测和跟踪,满足了实时性、连续性以及抗干扰能力强等要求,对于旋转及遮挡均具有较强的鲁棒性,为实现基于手势控制的人机交互提供了一种有效方法。

Human-Robot Interaction Implementation Based on Real-time Gesture Detection and Tracking

In order to realize automatic recognition and successive tracking of gesture, a gesture recognition and tracking algorithm is proposed. First, the gesture detector is trained off-line to achieve the automatic recognition . Secondly, an improved algorithm based on Shi-Tomisi is utilized to extract the reliable and stable feature points. Then, the feature points are tracked by the KLT tracker, when the feature points are tracked successfully, the gesture location is determined by solving the affine transformation matrix. When the object is occluded or in large scale rotation, the feature points are disappeared, the Kalman filter is added into the KLT tracker to estimate the location of gesture object, the successive tracking of gesture object is achieved. It estimates the area where the gesture may exist at the same time, which reduces the search range and improves the processing speed of the tracker. Finally, the algorithm is successfully used in a Human-Robot interaction system. Experimental results show that the tracking accuracy of the algorithm in simple background is 99.54% and the tracking accuracy in complex background is 98.24%. Experimental results indicate that the algorithm can detect and track the gesture object fast and accurately, it can satisfy the requirements of real time, continuity and strong anti-jamming, it also has strong robustness in occluded environment and large scale rotation, provides an effective method for gesture based Human-Robot interaction.