Research On Key Technologies Of Micro Hand Gesture Recognition Based On Millimeter Wave Radar

With the coming of the fifth-generation(5G)communication system,Internet of Thing(Io T)and artificial intelligence,various applications of Virtual Reality(VR)and smart wearable devices have emerged constantly.Human-Computer Interaction(HCI)plays a critical role in achieving seamless connectivity and promoting the integration between human society and cyberspace.Hand gestures are a natural form of human communication.Over the past few years,there has been a growing interest in the research field of hand gesture recognition.Micro hand gestures only involve the subtle movements of fingers.In contrast with coarse gestures,they can realize fine-grained interactive operations.Therefore,micro gesture recognition systems have been widely applied in HCI scenarios that require high identification accuracy,such as driver assistance systems and wearable devices.What’s more,among radio frequency(RF)based sensors,with the properties of high spatial resolution,high reliability,and easy integration,millimeter-wave radar-based sensors have shown promising performance in short-range micro hand gesture sensing.In this dissertation,a micro hand gesture recognition system is presented,which can be deployed on a commodity multiple input multiple output(MIMO)millimeter-wave radar platform as software,without any hardware modification.Three key issues are studied deeply in this dissertation,including a joint motion parameter estimation algorithm of finger targets,a clutter suppression method for micro hand gesture recognition,and gesture feature extraction and classification algorithms.The distinctive features of this dissertation are as follows:(1)Research on a joint range-angle estimation algorithm of finger moving targets.Firstly,to improve the resolution of the angle of arrival(AOA),a binary phase multiplexing based technique is utilized to realize the virtual array extension on the MIMO radar.Then,aiming to address the limitations of traditional joint range-angle estimation algorithms,such as low accuracy and high complexity,an Extrapolation-MUSIC based joint range-angle super-resolution estimation algorithm is proposed by combining an auto-regression model with the multiple signal classification algorithm.The experimental results based on actual measurement radar data show that the proposed algorithm can provide precise range-angle measurement and be used to track the motion trajectory of finger targets.(2)Research on a clutter suppression method for micro hand gesture recognition.In the process of performing gestures,the information of clutter targets is shown on range-angle maps.In terms of energy,velocity and AOA information,the motion parameters are similar between the clutter and real targets.Therefore,these clutter targets cannot be eliminated by traditional energy estimation based methods.By jointly taking feature differences in spatial position and energy into consideration,an innovative clutter suppression method based on an unsupervised convolutional auto-encoder network is proposed.The proposed method has the capability of reconstructing clean range-angle maps from noisy input according to the reconstructed cross-entropy minimization criteria.Finally,extensive experimental results based on both simulation and actual measurement radar data demonstrate that the proposed method can realize clutter suppression in the application of micro hand gesture recognition.(3)Research on feature extraction and classification algorithms.Due to the small range and slow velocity of finger movements,the spatial position difference between various gestures within a short period is not obvious.Using the obtained range-velocity and range-angle maps,a joint temporal-spatial feature extraction method is proposed by integrating the range,velocity and angle information.Then,to address the challenges of data pre-segmentation and the coupling relationship between gesture detection and classification,a recurrent convolutional neural network combined with a connectionist temporal classification algorithm is proposed to recognize micro hand gestures with high accuracy using continuous input radar data streams.The experimental results show that the proposed system is able to achieve high recognition rate of 96%,which is higher than state-of-the-art micro hand gesture recognition systems.In addition,the conclusion in this dissertation can be used for a real-time hand gesture recognition system design.