Research On Hand Gesture Recognition Method Based On Ensemble Classifier

The accuracy and anti-interference of gesture recognition are of great importance to improve human-computer interaction and promote the transformation of gesture recognition to the direction of ease of use and practicality.However,the complex background noise as well as small target and confusing gestures have become the technical bottleneck of gesture recognition in current practical use.To address the above problems,this thesis focuses on noise background underhand part cutting,feature fusion recognition network and integrated learning algorithm and other technologies,the main research content is as follows.(1)To address the problem of incomplete hand segmentation with noisy background,a hand threshold segmentation algorithm incorporating skin color information is proposed.The algorithm first uses the Cr channel image in the YCr Cb color space to extract the skin color sensitive region,then uses Gaussian filtering to filter out part of the background noise in the Cr channel image to reduce the effect of noise on the hand segmentation results,and finally uses the Otsu method to segment the hand in the grayscale image.The experimental results show that the proposed hand segmentation algorithm in this thesis is robust to different background noises and the hand segmentation results are more complete.(2)To address the problems of difficult extraction of small target gesture features and low recognition accuracy,we propose a convolutional neural network for hand gesture recognition by fusing multi-scale feature maps.First,the network extracts the shallow feature maps rich in information about small gesture targets and fuses them to the final feature maps used for classification to enhance the amount of information about small target gestures in the feature maps.Then,to avoid the problems of gradient explosion and slow convergence caused by the increase of parameters,a normalization layer is added to the network to perturb the data distribution and thus improve the network’s ability of fast convergence.The experimental results show that the convolutional neural network designed in this thesis improves the average accuracy of small-target gesture recognition by 4.2% compared with the pre-fusion network,and the network computation time does not increase significantly.(3)To address the problem of low accuracy of single classifier on confusing gestures,an integrated classifier for gesture recognition based on adaptive weights is proposed.First,support vector machine,Hu rectangular shape matching classifier and the feature fusion network designed in this thesis are selected as individual classifiers of the integrated classifier by experimental comparison.Then,we propose an adaptive weight assignment algorithm to increase the weight of individual classifiers with high probability of predicting categories in the integrated classifier,and provide a rejection prediction option to further improve the reliability of the integrated classifier.The experimental results show that the proposed adaptive weight-based integrated classifier for gesture recognition has a good ability to recognize confusing gestures,with an average recognition accuracy of 98.3% on 10 different gestures,while ensuring the real-time performance of the classifier.