HLAC-Based Feature Extraction Methodology And Its Application To Hand Gestures Recognition

Hand gesture recognition(HGR)is a hot topic in recent years.It has been widely used in human-machine interaction.Hand gesture recognition techniques play a critical role when applied to a natural and bare hand for depicting gestures in verbal communication.The main goal for hand gesture system is to attain the significant classification rate for the recognition of different hand gestures.In this dissertation,we are focusing on the Higher-order local autocorrelation(HLAC)feature and its extraction methodology to build feature model that is used to solve the problems of HGR system.Compared with conventional feature extraction techniques,the higher order local autocorrelation method has the capacity to restrict orders and spatial displacements of each pixel of hand image for gesture recognition.Hand gestures recognition is the most challenging task due to the following key issues.Firstly,we need to detect human hand which is very changeable and surrounded with complex background.Secondly,the high computational complexity of the existing techniques.Thirdly,the utility of HLAC in hand gesture recognition domain is limited to few features at a time.This is because of using traditional methods of exploiting features.This dissertation aims to improve performance and computational efficiency in the hand gesture recognition task.The main contributions and relevant achievement of this dissertation are;1.The first step in most hand gesture recognition is the hand region detection and segmentation.This segmentation can be a particularly challenging task when it comes to complex backgrounds and varying illuminations.In such environments,most hand detection techniques fail to obtain the exact region of hand shape especially in the cases of dynamic gestures.We employ a combination of existing techniques,based on motion detection and introduce a novel skin color classifier to improve segmentation accuracy.Motion detection is based on image differencing and background subtraction,and skin color detection is accomplished by using color classification technique.The experimental results indicate that the proposed method can efficaciously improve segmentation accuracy.2.The traditional higher-order local autocorrelation(HLAC)extraction method is more complex which needs the high computational cost to extract the desired number of features.Herein we improve HLAC extraction method into a low computational cost.The extracted HLAC features are inherently related to its local autocorrelation features and insensitive to shift-variance and scale variance.Thus,the limitations with respect to features distribution and size of the hand in an image are minimized.With this way,the ambiguity of computation has been reduced significantly.The experiment results show that HLAC features are effective for HGR with high recognition accuracy under low computational cost.3.To effectively exploit large numbers of features at low computational cost,a hierarchically staking low-order local autocorrelations feature extraction methodology is presented.This method is motivated by remarkable success in deep learning.The key concept of this technique is to decompose the computation of HLAC into multilayer network rather than computing everything into one input layer as the original HLAC does.Experimental results show that improved HLAC represents powerful discriminative information with a moderately sized feature vector compared to traditional one.4.We introduce a new feature extraction method based on the integration of HLAC and HOG features.We study the statistical relationship between the features in each training sample image,and we use this relationship to classify features.In order to evaluate the performance of the proposed method,we compared with other related methods such as Histogram of Colour(HOC),Histogram Oriented Gradients(HOG),Local Binary Pattern(LBP)and Gabor Filter method.Using artificial neural network,the empirical classification results show the optimal performance achieved by combining HLAC and HOG features.The proposed approach gives the best performance in terms of recognition accuracy(92.85%).The findings and conclusion proposed above will provide a new perspective for hand gesture recognition and will encourage promising investigations along the lines suggested.