Research On Dynamic Hand Gesture Recognition Algorithm Based On Convolutional Neural Network

With the popularization of smart terminals,human-computer interaction has become an important part of people's lives,and the introduction of more natural and convenient interaction methods has become an urgent need for people.Dynamic gesture recognition has attracted the attention of researchers due to its simple and natural characteristics.At present,dynamic gesture recognition faces difficulties such as flexible and diverse gesture actions,insufficient network real-time performance,and complex background interference.This thesis uses convolutional neural networks to study dynamic gesture recognition algorithms.The main work and innovations are as follows:Aiming at the existing uniform sampling method based on optical flow,the special case that the number of samples to be sampled is greater than the length of the subsegment is not considered.This thesis proposes a circular and uniform sampling method based on optical flow.First,the number of samples in each sub-segment is determined by calculating the ratio of the optical flow value of each sub-segment to the total value of the video optical flow.Then the normal sub-segment and the special sub-segment are uniformly sampled and cyclically re-sampled respectively,and finally a unified key frame sequence is obtained.Using different sampling methods according to different situations improves the robustness of the key frame extraction algorithm.In order to solve the problem of too many parameters of the three-dimensional convolutional neural network,the network efficiency is low.This thesis introduces the idea of splitting the twodimensional convolution kernel into the three-dimensional convolutional neural network,and proposes a method to split the three-dimensional convolution kernel into temporal convolution and spatial convolution to reduce the amount of network parameters.Then use this method to improve the three-dimensional Res Ne Xt network,and use the improved network for isolated dynamic gesture recognition.The experimental results show that the improved network can effectively improve the efficiency of the network while slightly reducing the accuracy.In this thesis,a hierarchical network structure is constructed and combined with a sliding window mechanism for continuous gesture recognition.The structure is divided into two levels: a gesture detection network and a gesture classification network.The first-level network determines whether there are gestures in the sliding window,and the second-level network classifies the video in the sliding window.Aiming at the problem that the classification network has multiple responses to a single gesture,a network state transition algorithm is proposed,which uses the previous state of the classification network and the number of consecutive undetected gestures by the detection network to control the state transition and working hours of the classification network.The network state transition algorithm solves the multiple responses problems of the classification network.At the same time,according to the characteristics of continuous gesture recognition and biological sequence comparison,this thesis introduces normalized edit distance as an evaluation index.The experimental results show that the average normalized edit distance of this method is 0.91,which is better than other existing methods,and it also has certain advantages in recognition rate.