Application Of Gesture Recognition Based On Capsule Network

With the development of economy and science and technology,people's demand for living conditions is getting higher and higher.In order to liberate inefficient productivity and improve social happiness,artificial intelligence emerges at the right moment.The development of computer science and intelligent algorithm is also contributing to artificial intelligence.Solving the related problems of unmanned driving has become a hot research topic.This paper mainly identifies and judges the gesture of the traffic police.The traditional use of professional wearables and identification training of the device parameters can solve this problem.However,the need for wearables of the traffic police and vehicle installation equipment is not good for the traffic police to fulfill their duties and the cost is high.It is an efficient and low-cost method to use image recognition to solve the problem of traffic police gesture.Convolutional neural network is generally used for image recognition,but it requires a large number of samples,long training time and high equipment requirements,and the translation invariance of convolutional neural network is not sensitive to the position relationship of traffic police gesture.In order to solve the neural network to carry on the deep study on the demand of data samples,long training time problem,the paper studies the capsule and the and the network Resnet18 neural network model is built on traffic signs and comparison,compared with Resnet18 network found capsule volume of learning on the sample data is low,the training of high precision,strong portability.In real life,the pictures taken by cars are not only the traffic police.For the pictures taken by cars,we should first identify the objects in the pictures,find the people in the pictures and judge whether there is a traffic police in the pictures,and then identify whether the traffic police are making gestures,and then judge the meaning of the gestures,and finally make decisions.The paper uses Python crawler to obtain relevant pictures,uses improved SSD model to identify people in the pictures and intercepts them,uses Alex Net network to train the dicclassification model for the recognition and judgment of traffic police pictures,because the data amount of intercepted pictures after crawling is too small to be used for model training,and then takes 1200 pictures by itself.The command gesture pictures of traffic police were carried out Resnet18 transfer learning,and the gesture recognition model was trained.Finally,the capsule network model was used to learn and analyze the gesture pictures of traffic police.The results show that under the same equipment and conditions and the same data set,the model accuracy of capsule network training is about 85%,much higher than that of Resnet18 network 60%,indicating that capsule network has a high accuracy in small samples and can preserve the posture information of the object in the image well,so it can achieve a good effect in the field of gesture recognition.