Human Lower Limb Gait Sensing Signal Acquisition And Gait Recognition

Gait perception and recognition is an important leading technology for wearable robots,which is an important basis for achieving human-machine coupling tasks.It has far-reaching application prospects and research value to carry out related research on human gait.At present,most of the tasks of human gait recognition based on gait sensing data are focused on the recognition of motion modes or different gait phases under a specific motion mode.In order to make wearable robots complete human-machine coupling tasks better and adapt to more complex human behaviors,it is very necessary to design a more fine-grained and high-precision gait recognition algorithm that can distinguish different gait phases under similar motion modes and ensure recognition performance during frequent gait switching.In response to the above needs,this thesis designs and implements a new Lightweighted gait acquisition system that can simultaneously collect three types of different sensing signals: surface electromyography of the human thigh,foot pressure of the lower limb,and knee/hip joint motion.Using this system to carry out data acquisition experiments,construct a gait recognition dataset,and carry out research work on multimode multi-gait cycle(29 classes)classification and recognition algorithm based on short-term gait sensing data samples(0.1s),including:(1)Different processing methods are used for preprocessing the three types of collected gait sensing signals.Discrete wavelet transform is used for denoising and DC removal for surface electromyography signals;Gaussian filter is used for smoothing foot pressure signals;IMU signals are repaired for anomalies and data restoration,and corresponding verification algorithms are designed to verify the processed data to ensure its rationality.(2)Considering that the recognition network in this thesis needs to be able to distinguish the difference of the same gait phase in similar motion modes,for example,it needs to be able to distinguish the difference between the left foot in front of the support state in flat walking mode,uphill mode and downhill mode,this thesis selects the feature set based on the prior knowledge of human gait for distinguishing different gait,and at the same time,the data collected by the gait acquisition system is retained.(3)In the gait recognition network,the gait sensing data is continuous time series data,this thesis adopts LSTM as the basic structure of the gait recognition network and incorporates the Attention mechanism into it,so that the gait recognition network designed in this thesis can quickly notice the key features in the gait sensing data,and then can achieve higher accuracy recognition effect by only relying on a small amount of gait sensing data.Finally,the recognition rate is 92%.(4)Several experiments were designed to validate the algorithm.The network structure of this thesis is compared with traditional FC networks and CNN,and is compared with cutting-edge "VWI-DNN" and "Stacked LSTM" gait recognition networks.The average number of false positives and the false positive rate of gait switching are also proposed to further validate the ability to cope with this difficult recognition problem in the process of gait switching.