| Sign language recognition(SLR)is a multidisciplinary research topic in pattern recognition.Because the continuous frames in sign language videos contain a lot of redundant information,extracting key frames to eliminate irrelevant information has always been the research focus in data preprocessing of sign language videos.In recent years,skeletal data has emerged as a new type of data in sign language research and has attracted more and more attention,which promotes the research on feature extraction of skeletal data.Meanwhile,with the penetration of deep learning methods into computer vision,building a high-performance network for sign language recognition is also an important research topic.This paper will study several key topics in sign language recognition.The main contents are as follows:We do optimization on traditional keyframe sampling algorithm.We design the optimized algorithm based on the idea that Dynamic Time Warping(DTW)distance can quantify the similarity between two sequences with different lengths.Besides,we do research on parameter settings of keyframe sampling algorithm,and provide a set of standardized operation for it.For the task of sign language recognition using skeletal data,we design a recognition system.Combined with the idea of Multi-View and Graph Convolutional Network(GCN),we design a feature extractor.This extractor uses GCN unit to get graph features.The graphs are generated by projecting the skeletal joints onto three orthogonal planes.During training and recognition,we design a two-dimensinal parallel architecture based on the attentional mechanism and the idea of feature fusion.The network can extract the spatial and temporal features of skeletal data and then fuse them for training and recogntion.Finally,we describe the whole architecture for the recognition system,which provides a model for the sign language recognition system using skeletal data as input. |
