摘要： 手势识别作为一种自然和谐的人机交互方式，具有广泛的应用前景，而传统手势识别方法准确率不高、实时性较差。为此，在DSSD网络模型的基础上，提出一种静态手势实时识别方法。自制手势数据集，通过K-means算法及手肘法选取先验框的宽高比，采用迁移学习的方法解决数据量小导致的检测精度低的问题，同时根据识别精度选择ResNet101为DSSD模型的基础网络，经DSSD模型的反卷积模块融合各个特征提取层的语义信息，加强对小手势目标的检测能力。实验结果表明，该方法识别静态手势的识别率达到95.6%，较基于Faster R-CNN、YOLO和SSD的手势识别方法分别提高了3.6%、4.5%及2.3%，其检测速度为8 frame/s，能够满足实时检测要求。

关键词: 手势识别, DSSD检测算法, K-means算法, 迁移学习, 小手势目标

Abstract: Gesture recognition,as a natural and harmonious way of human-computer interaction,has wide application prospects.Aiming at the problems of low accuracy and poor real-time performance of traditional gesture recognition methods,this paper proposes a real-time recognition method for static gestures based on DSSD network model.In this paper,a gesture dataset is created and the aspect ratio of the prior box is selected by the K-means algorithm and the elbow method.The transfer learning is used to solve the problem of low detection accuracy caused by small dataset.At the same time,the ResNet101 is selected as the basic network of the DSSD model according to the recognition accuracy.Then,the deconvolution module of the DSSD model fuses the semantic information of each feature extraction layer to enhance the detection ability of small gesture targets.Experimental results show that the recognition rate of the static gesture recognition of this method reaches 95.6%,which is 3.6%,4.5%,and 2.3% higher than those of the gesture recognition methods based on Faster R-CNN,YOLO,and SSD.Moreover,the detection speed of this algorithm is 8 frame/s,which can meet the requirements of real-time detections.

Key words: gesture recognition, DSSD detection algorithm, K-means algorithm, rransfer learning, small gesture target

中图分类号: 

• TP391