基于数理统计特征的人机交互图像手势识别

Gesture Recognition of Human-Computer Interaction Image Based on Mathematical Statistics

在人机交互领域中手势识别的应用前景十分广阔,在诸多领域中均为人类带来了无限便利。基于数理统计特征设计一种人机交互图像手势识别方法,实现了基于数理统计特征的手势信息获取与基于识别结果的人机交互。对人机交互图像实施图像灰度化处理、二值化处理、平滑处理、边缘检测和轮廓提取处理等一系列预处理。通过OpenCV提取二值化人机交互图七个具有尺度不变性、旋转不变性、平移不变性的Hu矩,前四个矩描述手势的图像椭圆、主轴方向角、面积、旋转半径这四个物理量,后三个矩描述的是图像对称性、重心、中心距。基于Darknet-19改进YOLO-V2网络的骨干网络,使模型能够预测的anchor box数量达到16×16×N个。基于改进YOLO-V2网络设计手势识别模型,模型的输入为人机交互图、提取的手势轮廓与Hu矩,实现交互图像手势识别。测试结果表明,设计方法在室内和室外的手势识别准确率均较高,通过该方法的手势识别结果能够实现人机交互。

The application prospect of gesture recognition in the field of human-computer interaction has brought infinite convenience to human beings in many fields.Based on mathematical statistics features,a method of gesture recognition for human-computer interaction images is designed,which realizes gesture information acquisition based on mathematical statistics features and human-computer interaction based on recognition results.A series of preprocessing such as image gray-scale processing,binarization processing,smoothing processing,edge detection and contour extraction are implemented for human-computer interactive images.Seven Hu moments with scale invariance,rotation invariance,and translation invariance are extracted from binary human-computer interaction images using OpenCV.The first four moments describe the physical quantities of the gesture′s image ellipse,principal axis direction angle,area,and rotation radius,while the last three moments describe image symmetry,center of gravity,and center distance.The backbone network of YOLO-V2 network is improved based on Darknet-19,so that the number of anchor boxes that can be predicted by the model reaches 16×16×N pieces.The gesture recognition model is designed based on the improved YOLO-V2 network.The input of the model is the human-computer interaction graph,the extracted gesture contour and Hu moment,and the interactive image gesture recognition is realized.The test results show that the design method has a high accuracy of gesture recognition both indoors and outdoors,and the human-computer interaction can be realized through the gesture recognition results of this method.