Research On Sign Language Recognition Based On Depth Information

The research of sign language word recognition aims to converts the dynamic video obtained into text information.The traditional sign language recognition algorithm often uses RGB camera to collect sign language information.With the development of industrial technology,Capturing deep information is easier and less susceptible to environmental light.Therefore,this paper uses the depth information to carry sign language words recognition that obtained by depth sensor.In 2015,the Xilin Chen team use the SO-SMP(Sparse Observation-Stable Marriage Problem)algorithm to score in the chalearn sign language matching game and perform better than the algorithm of 2013 chalearn champion,but there was only problem about the selection of K-value in k-means.The problem is that they haven't considered the Sum(Sum Of The Squared Errors)value,So the accuracy is not high on the non-specific dataset,this paper improves on the basis of the SO-SMP algorithm and improves the gesture recognition accuracy.The main work includes:(1)A static sign language database is established,it includs 40 different kinds of china sign language words,then author using the Gaussian skin color model to segment the hand shape for obtaining the HOG(Histogram of Oriented Gardient).Then The HOG is used to get feature template.Finally,KNN(K-Nearest)was used.it get the 100% recognition rate when the training data is sufficient.(2)An improved SO-SMP algorithm is proposed.Firstly,the data is dealed refinedly,and black hole region of the depth data is filled by median,bilateral and joint bilateral filter,and The experimental results on the set are analyzed.The K value is re-evaluated by the combination of the contour coefficient and the elbow method.The accurate K value is used to redefine the exact template class in the template matching process,which reduces the search space of the template state class and improves the matching of the two sign language state sequences.Accuracy,and which is higher than the original algorithm recognition rate on the specific person's sign language data set.(3)For the problem that the accuracy improvement of the non-specific sign language sign language data set is not significant,the normalization operation of the human body in the chalearn data set is performed to improve the recognition accuracy of the sign language on the non-specific person's sign language data set,and then consider the HOG feature itself is too large in dimension,and it is too time-consuming in the matching process.It constructs a comprehensive feature combining Markov distance,Euclidean distance and dip angle between joints and reduces the dimension of the feature,although there is a recognition accuracy of dynamic sign language words.The reduction reduces the time-consuming of the algorithm to about one-tenth of the original algorithm.