Fusing Multi-feature Dynamic Gesture Recognition Based On Multiple Kernel Learning

With the increasing popularity of computers in today's society,the study of human-computer interaction will have a positive impact on it.Among them gestures are considered to be a more natural,creative and intuitive human-computer interaction technology.With the advent of depth cameras such as Kinect,depth data has become an important means of gesture recognition research.In the current stage,the traditional feature extraction methods have achieved satisfactory results in dynamic gesture recognition.However,we ignore the role of the different structural features in the gesture as well as the non-linear relationship.The main work of this paper is to fuse the heterogeneous features and construct the model through the representation of nonlinear kernel to accurately classify the dynamic gestures.In order to obtain effective feature information from dynamic gestures and explore the nonlinear relationship between underlying features,this paper designs a multi-feature fusion based dynamic gesture recognition method based on multi-kernel learning.In order to obtain highly discriminating feature information,the depth motion diagram of gestures is considered as the source of feature information.Spatial multi-scale binarized histograms and gradient histogram features are extracted as three-dimensional shape structure information of gestures.Then temporal-spatial gradients of gestures are extracted.The histogram feature maps temporal features to the frequency domain using Fourier transforms in time series.This way,the feature vectors can be aligned on the one hand and the feature presentation mode can be changed on the other hand.Finally,the two feature features are merged to obtain the final feature descriptor.Defined as PDL HF~2.Considering the feature dimension of the feature descriptor is too high,it is not easy to find the linear relationship in the low-dimensional space and reduce the generalization ability of the classifier.This paper proposes a multi-core learning method to select the appropriate feature space training classification model.Firstly,we use the regularization method to preprocess the feature descriptors.Then we use different kernel inner product relations to redefine the two feature vectors as the elements of the new feature matrix.We use the kernel alignment algorithm to calculate the weights of different kernels and obtain a new one.The weighted fusion kernel matrix is??used to apply the new feature matrix to the data set for feature selection of each gesture.Finally,the performance evaluation of the experiments on the two gesture datasets MSRGesture3D and SKIG can obtain 100%and 99.44%recognition rates,respectively.The experimental results show that the proposed method is more optimal than the current method.In summary,we use the public databases MSRGesture3D and SKIG to obtain the feature data using a series of spatio-temporal structure feature extraction algorithms,and use the kernel learning method to map the feature matrix and compare the results.The results show that the proposed method is robust to noise and very effective for the classification of hand gestures in depth data.It also proves that the proposed method has a better classification effect than the traditional feature extraction algorithm.