The Research On Gait Analysis Based On Dynamic Vision Sensor

Gait characteristics,as the characteristics of human walking,contain a large amount of unique information about individuals,such as height,shape,and length of limbs,as well as cadence,stride,and posture during walking.At the same time,the collection of gait information does not require participants to actively join,and can achieve the identification and authentication of their identities without notifying participants.The gait analysis research based on the traditional RGB camera requires complex calculations such as background removal.Therefore,this paper mainly proposes a new type of bionic sensor: dynamic vision sensor,based on which to perform gait analysis and research.Compared with traditional RGB cameras,dynamic vision sensors have the advantages of low latency,low power consumption,high sampling frequency,and wide dynamic range.At the same time,the output of the dynamic vision sensor is an asynchronous event sequence,which describes the changes in the brightness of the pixels in the sensor's field of view.Only the area where the brightness changes is retained to achieve the effect of automatic background removal.However,existing computer vision algorithms based on frame images cannot be applied to process the event streams directly.In order to solve this problem,this paper proposes a gait recognition model based on convolutional neural networks.By converting the asynchronous event stream collected by the dynamic vision sensor into an "event" picture,similar to the heat map in thermodynamics,and applying a convolutional neural network to build a gait recognition model.In addition,for the noise events generated by the sensor,this paper innovatively proposes a noise removal algorithm that uses the speed of optical flow,which is different from other neighborhoods to filter the noise events based on the occurrence of temporal and spatial events in the neighborhood.The difference between events is to identify whether it is noise by calculating the optical flow velocity of the event stream,and designed corresponding experiments to verify.Two qualitative and quantitative experiments prove that the noise reduction algorithm proposed in this paper is better than other current noise reduction algorithms.Finally,in order to verify the performance of the gait recognition model based on the convolutional neural network,two gait event stream data sets were produced in this paper:DVS128-GAIT and EV-CASIA-B.DVS128-GAIT is collected under actual environment,and EV-CASIA-B is converted from the published gait data set CASIA-B.The experiment proves that the gait recognition model proposed in this paper applied to dynamic vision sensors can achieve an accuracy of about 96% in the data set collected under real conditions.After the converted data set is verified,its recognition accuracy is comparable to that of traditional RGB cameras.