Research On Gesture Tracking And Recognition Based On Ultrasound

In the coming era of Internet of Everything,human-computer interaction is an crucial part between humans and things.Human-computer interaction based on gestures is very convenient,and how to sense gestures is also one of the hot research topics.At present,the research on gestures sensing is mainly based on optical signals and electromagnetic signals.There are few researches on ultrasound based gesture recognition,however,ultrasound has the advantages of environmental adaptability,low power comsumption and low cost.This technology has potential application in wearable devices and smart house.Therefore,this thesis studies the gesture tracking and recognition based on ultrasound.The main works of this thesis are as follows:For application in two-dimensional gesture trajectory recognition,the sensing principle is explained and the hardware structure is constructed.The transmitting and reception of ultrasonic signal are carefully designed;besides,chips,circuits and other parts are introduced.Ultrasonic transmission and reception system are made and exhibited.The design of the transmitting circuit and the echo amplifying circuit is verified by checking waveforms of transmitting signal and reception signal.In order to eliminate the signal noise after echo acquisition,three filtering algorithms are compared and the echo signals are filtered.The echo of environment and hand is analyzed,and an echo automatic calibration algorithm based on crosscorrelation operation is proposed.To solve the problem that the distance information is hardly be extracted from echo due to distance attenuation,a distance information extraction method based on dynamic threshold and variable gain is proposed.According to the geometric relationship of the transmitting and receiving transducer,the elliptical positioning model is established and the influence of geometric parameters on the positioning accuracy is studied.In terms of trajectory processing,the differential-based start and end point detection algorithm and the trajectory outlier processing method based on component difference and hopping detection are proposed.What is more,a two-dimensional motion experiment system is set up to verify the practicability of the algorithm.To recognize the trajectories of Arabic numerals and uppercase letters,a data set for training and testing is constructed.The random forests model is selected as classifier,and combined with the actual situation of gesture trajectories.The trajectory feature extraction methods based on dynamic grid and direction angle are used respectively,and then the two feature sequences are merged.The recognition results of the two feature extraction methods used alone and mixed are compared.For the better recognition effection,the influence of model parameters on the results is studied.The final experimental results illustrate that the processing of the trajectory,the feature extraction method and the random forests model can recognize trajectories well,and the accuracy is over 89%.