Research On Gesture Recognition And Human Behavior Recognition Algorithms Based On Deep Learning

With the development of artificial intelligence,gesture recognition and human behavior recognition have become an important research topic in the field of artificial intelligence.In the field of intelligent transportation,gesture recognition can realize vehicle-mounted human-computer interaction,and human behavior recognition can realize abnormal detection of driver's driving behavior.Gestures are one of the ways to convey information between people,and also an important way to communicate between deaf and mute people and normal people.Therefore,understanding the meaning of sign language is very important.Using artificial intelligence technology to recognize sign language automatically will improve the efficiency of interaction between deaf-mutes and the outside world to a certain extent,and solve the communication difficulties of deaf-mutes.Human behavior is an important yardstick to measure the change of human posture.Analyzing the change of human posture from the data acquired by sensors can judge whether the elderly fall or not.The first time to send out rescue information,start the GPS positioning system,and provide the possibility of emergency rescue.Based on these two research backgrounds,this paper carries out in-depth learning technology research on gesture recognition and human behavior recognition.Based on the research status of gesture recognition and human behavior and the elaboration of related deep learning theory,this paper analyses gesture recognition algorithm and human behavior recognition algorithm,and carries out relevant research from two aspects of gesture recognition and human behavior recognition.In the aspect of gesture recognition,10 kinds of gestures commonly used in human-computer interaction are collected to establish a database,and gesture recognition is carried out based on YOLOv3,which achieves 92.6%of the mAP on the test set.An improved convolutional neural network is proposed.The mAP on the test set is finally achieved 98.2%by using data enhancement technology,Maxout activation function and increasing the the input image,Which improves the recognition rate of gesture.A lightweight convolution network based on MobileNet is introduced to compress the model.The compressed model is only about 10M,which reduces the memory occupied by the model.The mAP of the test set based on the compressed convolution network is 97.1%.At the same time,the accuracy of gesture recognition and the memory occupied by the model are taken into account to meet the real-time requirements.In human behavior recognition,this paper designs a convolutional neural network to recognize human behavior.Eight different types of human behavior data are extracted from MobiFall,MobiFall and SisFall datasets.Kalman filter technology is used to process the three-axis acceleration data and the three-axis gyroscope data acquired by sensors and transform them into noise reduction data.The RGB image pixel value is trained by the convolution neural network designed in this paper,the effect is compared with that of the convolution network designed by KNN algorithm.The test set is evaluated.The average accuracy of using KNN algorithm and designing the convolutional neural network was 90.06%and 97.42%respectively.