Gait Feature Analysis And Gait Pattern Recongnition Based On ANN

Gait refers to the people’s posture when walking, a basic biometric. Gait has theunique advantages of non-invasive, difficult to conceal, low distant requirement. Gaitcan reflect the problems in human physiology,motion control,and psychological,somuch attention to gait is paid in recent years.Research institutes and universities athome and abroad has done much research on the area. In the gait feature extractionand gait recongnition field, we still need further deep study.Although the research of human gait is mainly concentrat in the camera visionand image precessing area, the invention of acceleration sensor provides a brand newmentality which makes the gait acceleration signal be an important part of gaitinformation. Using acceleration sensor the convenient system can acquire real-timegait data which has less environmental interference factors. So the thesis presents gaitfeature analysis and gait pattern recongnition based on3D kinematic accelerationsignal.The thesis designs a gait acceleration data acquisition system consisted of thetriaxial acceleration sensor MMA7260Q, STM32F103a cortex_M3seriesmicrocontroller and a W25Q128BV flash memory which are respectively as thesystem’s data acquisition part, data processing part and data storage part.Formulatethe experiment rules,the experimental objects wear the data-acquisition sysytem to getgait signals in different conditions.By this method,we establish a gait feature databaseand a gait recongnition database.In order to ensure the gait signal’s accuracy and regularity,some preprocessingmethods,such as location correction,denoising, normalization are applied to theorignal acquisited gait signal.In the data type transformation field, the gait data in“212” form should be converted to the decimal form; We use geometric way to revise the gait data to the standard coordinate system; In order to denoising gaitsignals,the thesis adopts “db8” wavelet basis to decomposition the gait signals;In thedata-normalization field, we apply the linear normalization way to scale the data tothe0~1area. The gait frequency algorithm mainly apply time windows division andthe method of setting spike threshold.By amplitude and peak interval limited,thealgorithm eliminates the flase step and improve the accuracy of calculation gaitfrequency.In terms of gait feature extraction,the gait signal is divided byzero_point,we can obtained some important information such as the gait stance andswing phase. The thesis uses phase symmetry index as an important standard fordetermining gait symmetry, distinguishs the subjects’different gait symmetry.Compare different nonlinear classification algorithms, neural network has thefollowing advantages:firstly,neural network has strong robustness and fault tolerance;secondly,nerual network has parallel processing structure,and its each unit can dealsimilar process at the same time,and the computation time is short;thirdly,neraulnetwork has strong adaptability,neurons connect each other in a wide variety ways.coupling strength between neurons in each layer has a certain plasticity, make it dealwith uncertain or unknown systems;Fourthly, nerual network can learn complexnonlinear relationship.So the thesis chooses neural network to build a classifier forgait aceeleratin signal with different states.The thesis mainly distinguish two modes which are dynamic gait pattern andstatic gait pattern,Particularly all the gait patterns include six static dynamic whichare sitting pattern,standing pattern,quickly walking pattern,slowly walkingpattern,upstairs pattern and downstairs pattern. According the summary to concretementality, the thesis adopts a staged method to reconigize different gait patterns. Inthe pre-classify stage,Signal Margin Area(SMA) and Average Energy(AE) are usedas the key feature to distingguish static and dynamic gait. they will be serve as theinput characteristics for the first neural network classifier.Then in the specificreconnition stage,in order to pevent BP nerual network over fitting,we apply MeanImpact Value(MIV) to compressing recongnition static and dynamic gait Features’sdimensions.The compressed features are the input value of neural network.We use the selected training gait data to train BP nerual network,then we use the trained neuralnetwork to recongnize different real time gait signals,the gait recognition accuracyis satisfied.Finally,I summarize the thesis,and come up with the improvement plan.