基于自适应多分类器融合的手势识别

为了提高基于可穿戴设备手势识别的性能,针对单分类器在手势识别时会出现偏向性的问题,提出了基于自适应多分类器融合的手势识别方法(Self-adaptive Multi-classifiers Fusion,SAMCF)。首先,针对统计特征无法表征复杂手势之间类内变异性和相似性的问题,SAMCF使用卷积神经网络(Convolutional Neural Network,CNN)自动提取具有强表征能力的深度特征;然后,采用多个基本分类器对提取的特征向量进行识别,并通过自适应融合算法决策出最优识别结果,解决了单分类器的偏向性问题;最后,基于数据手套采集的数据集,验证了模型的鲁棒性和泛化能力。实验结果表明,SAMCF能够有效地提取手势的深度特征,解决单分类器的偏向性问题,提高了手势识别的效率,增强了手势识别的性能,对字符级手势(美国手语和阿拉伯数字)识别的准确率达到98.23%,较其他算法平均提高了5%;对单词级手势(中国手语)识别的准确率达到97.81%,较其他算法平均提高了4%。

Hand Gesture Recognition Based on Self-adaptive Multi-classifiers Fusion

In order to improve the performance of hand gesture recognition based on wearable devices,a hand gesture recognition method(SAMCF)based on self-adaptive multi-classifiers fusion is proposed to solve the bias of single classifier in hand gesture recognition.First,for statistical features that cannot characterize intra-class variability and similarity between complex gestures,SAMCF uses a convolutional neural network(CNN)to automatically extract depth features with strong representation capabilities.Then,SAMCF uses multiple basic classifiers to recognize the extracted feature vectors,and determines the optimal recognition result through self-adaptive fusion algorithm,which solves the bias of single classifier.After that,the robustness and genera-lization ability of the model are verified by using the data set collected by data glove.The experimental results show that SAMCF can effectively extract the depth features of gesture,solve the bias of single classifier,and improve the efficiency of hand gesture recognition and enhance the performance of hand gesture recognition.The recognition accuracy of character level hand gesture(American Sign Language and Arabic numerals)is 98.23%,which is 5%higher than other algorithms on average;the recognition accuracy of word level gesture(Chinese Sign Language)is 97.81%,which is 4%higher than other algorithm on average.