Research On Static Gesture Recognition Method Based On SAE Multiple Features Fusion

With the rapid development of computer vision and artificial intelligence technology,machine-centric interaction is gradually being replaced by human-centered interaction.The society urgently needs a more human ized and intelligent interaction mode,in which gesture interaction is intuitive.The characteristics of convenience,naturalness and so on make it an important way of interaction.The current gesture interaction is still at the primary level.It is very important to study static gesture recognition based on computer vision.In order to improve the recognition accuracy and recognition speed of gesture recognition,this paper designs and implements a static gesture recognition system based on computer vision to complete the recognition of predefined static gestures.The system mainly includes gesture data processing,gesture image feature extraction and classification.Identify three parts.Firstly,the classical feature extraction algorithm,dimension reduc tion algorithm and classification algorithm are discussed.The principles and implementation steps of HOG algorithm,LBP algorithm,PCA algorithm and SVM algorithm are introduced in detail.The advantages and disadvantages of each algorithm are compared experimentally.The features extracted by a single classical algorithm have higher dimensions and features.The category information that does not contain the gesture image affects the classification speed and accuracy.To solve this problem,this paper adopts an improved algorithm and introduces a sparse self-encoder(SAE)to reduce the dimensionality of multiple features.Features not only have lower dimensions,but also overcome the limitations of single feature rendering ability.SAE is an unsupervised deep learning network that requires a large amount of data for learning training.In order to overcome the problem of less training data,this paper is in the original data set.Based on the expansion,the expansion methods are rotation,symmetry and translation.Experiments on the JTD gesture database prove that using new features can effectively improve system performance.Secondly,the relationship between SVM classifier and features is studied in depth.The feature extracted by SAE algorithm lacks the inf ormation of gesture image itself,which leads to the increase of the misclassification rate of the same class of samples.To solve this problem,this paper proposes an integrated classifier algorithm based on SVM.Two different classifiers KNN and XGBoost are added,and the particle swarm optimization algorithm is used to select the integrated parameter with high diversity and high precision to construct the integrated classifier to classify the features.The experimental proof of the JTD gesture database i s presented in this paper.The method achieved a classification accuracy of 96.67%,and the accuracy was improved by 2.53% compared to the single classifier.Finally,the SAE-(HOG+LBP)-SVMs static gesture recognition system based on Microsoft Visual Studio 2015 and OpenCV environment is designed and implemented.The system mainly includes image preprocessing module,SAE training module and classifier module.The JTD data set is used.The expansion is made into a training set,and the gesture images are collected into a test set under normal lighting conditions.The experiments on the JTD test set and the self-made test set prove that the system has higher accuracy and faster than the classical recognition system.Identifying speed,the improved algorithm is feasible.