Study On Key Technologies For Ambulation Behavior Perception Based On Plantar Pressure Distribution

Behavior perception based on artificial intelligence can considerably extend the capability of human itself in processing and utilizing behavior information of surrounding environment. Behavior perception research can push forward the development of intelligent environment and personalized service, and it can also propel the monitoring and warning of public safety to be more intelligent and automatic. Behavior perception research is a multidisciplinary area, which involves data mining, machine learning, computer vision and sports biomechanics. It is a quite wide and complex research area, which has attracts numerous scholars in recent years.According to different methods of information acquisition, behavior perception can be divided into two categories:method by visual information and method by non-visual information. The visual method has got widely developed because of the popularization of different image sensing devices. However, visual behavior perception system is susceptible to many different factors, such as occlusion, lighting condition, view angle and distance; furthermore visual method may cause privacy invasion problem. On the other hand, non-visual method is a good complement to visual method. In recent years, different inertial and magnetic sensors have been introduced into the area of behavior perception, however, one major drawback for those sensors is that they have to be installed on the body of human beings, which may cause some degree of uncomfortable feelings.This thesis starts from the construction of a non-visual, non-invasive and day-and-night behavior perception system. In this study, ambulation behavior was taken as major research objective; a large flexible force sensitive floor system was chosen as main carrier of behavior information; key elements in behavior perception system such as object segmentation, tracking and recognition were deemed as main themes. More specifically, the following sub-subjects were explored in this study:Firstly, investigation of how to sample, analyze and segment footprints for plantar pressure image. The plantar pressure sampling system based on flexible force sensitive floor has the characteristics of high resolution, high sampling rate and large area. Therefore, a module-based method is developed to build a network to solve data transfer and area-extension problem. Single plantar pressure image is composed of instant pressure value of every pressure sensitive sensors in the array, it has gaussian and impulse noises just existed in other kinds of images and also have a kind of time-relevance noise which are caused by some kind of manufacturing problems of flexible sensors and little distance between sensor elements. Also high sampling rate determines that the execution time allowed for denoise algorithm is very short. Therefore in this study, starting from a previous algorithm, a switch-median algorithm was adopted to perform prefiltering of noise, and then a two-dimension FIFO technique is utilized to realize fast execution of average filtering in time-relevant three-dimension filter windows. By such measures, the overall execution time of denoise algorithm has been reduced to a great extent. Footprint segmentation is the prerequisite of ambulation behavior perception research, to address the discrete distribution feature of data points, a density-based algorithm is utilized to realize the division of different data points into different groups, and then under the constraint of footprint shape descriptors, different data groups are combined to realize the segmentation of different footprints.Secondly, study of gait recognition based on plantar pressure distribution. As an important subsystem of behavior perception, gait recognition based on plantar kinetics has the characteristics of good concealment and long action distance. Therefore it has been developed quickly in recent years. As a biomechanic finding, plantar pressure distribution demonstrates somewhat personalized localization feature. According to this kind of phenomenon, a novel gait recognition method is proposed. In this method, several key points on the hump-shape GRF curve are firstly extracted and then the plantar pressure image corresponding to the moment of such points are utilized to construct spatio-temporal HOG feature, which not only describes the localization feature of plantar pressure distribution, but also embodies the time-variant feature of GRF curve. Based on such STHOG feature, a RBF-kernel based SVM algorithm was applied to realize the gait recognition in the end.Thirdly, research of a multi-footprint tracking algorithm. Based on the markov feature of ambulation behavior of human beings, the particle filter which has good tracking performance for non-linear and non-gausian signal has been adopted to be the main tracking framework in this study. At the same time, according to the basic feature of ambulation behavior, an action mathematical model for state prediction is proposed and for specific actions such as stop, turn back, a separate model is built for it correspondingly. Then during tracking process, model switching is performed based on observing information. Finally, footprints generated during segmentation stage were used as observation state, and footprint tracking is realized under the guidance of bayesian decision rule.Finally, study of detection algorithm for abnormal gait based on gait kinematics. Ambulation behavior of human beings is a complex and accurate process, which is determined by the dynamic interaction of central nervous system and feedback system of muscle. Many neurogenic and age-related diseases can cause such process to be problematic and a direct demonstration of such problems is abnormal gait. This study took ambulation behavior perception under intelligent dwelling house as the prerequisite to realize the daily gait monitoring of general public. The proposed algorithm utilized HMM to represent gait kinematics, and the gait classification is realized by HMM similarity measurement.In summary, this thesis takes the construction of intelligent monitoring and human-computer interaction as research direction. To address several key components of behavior perception such as object segmentation, tracking and recognition, some useful exploration and jobs are performed based on plantar pressure distribution information. We hope our job can be a beneficial hint and reference to a wide and deep exploration of research area for ambulation behavior perception.