Research On Human Activity Information Acquisition And Recognition

Human activity owns features of perceptibility, noninvasive and stability; it can reflect people's intention. Human activity recognition is currently one of the most active research topics in the field of biometrics research; it has potential application prospects in pervasive computing, virtual reality, sport training and health care. In recent years, although human activity recognition with wearable sensors received increasing attention, it only relies at the stage of theoretical exploration; many theoretical and technical problems remain open. Human activity information can be captured by inertial sensors such as accelerometer, after feature extraction and modeling, human activities can be recognized through statistics and machine learning algorithm. How to map low-level sensor data to higher level abstractions is the key to activity recognition. This paper focuses mainly on the method of acquisition rich data about human activities, gait recognition based on gait acceleration, automatic feature extraction, short-term activity recognition and high-level human activity recognition.The diversity of environment and the complexity of human activity affect the information extraction. A portable accelerometer-based system which used to accurately record human activities is designed, and the method of motion detection from sensor node, sensor data calibration and denoising are also analyzed.Aiming at the means of gait parameters extracting based on the portable gait data acquisition system, the method of measuring cadence, step and stride lengths by an autocorrelation procedure is presented, and gait symmetry could be estimated by analysis the ratio of correlation coefficient between signal step and gait cycle. The relation between walking speed and cadence, step and stride lengths is also given. This thesis proposes a kind of nearest neighbor algorithm which used to identify human individual through features of gait acceleration data in time domain and frequency domain. Experiments demonstrate the effectiveness of our algorithm.For the similarity and instability of short-term human activity signal, Discrete Hidden Markov Models used for gesture recognition are established with the combination of feature extraction by k-means clustering. The k-means clustering algorithm selects a small number of critical features from a large set by ranking different features according to the quality of the resulting clustering, and the small feature subset are taken as the input of DHMM, so as to identify different gestures. Experiments show that the feature selection algorithm can not only reduce the complexity of Discrete Hidden Markov Models, but also improve the recognition performance, and the method proposed in this thesis can be used for different kinds of sensor data, and extended to other kinds of short-term human activity identification.Usually the recognition of daily human activity need to collect multi-sensor data, this thesis explores an information fusion algorithm based Na(?)ve Bayes so as to obtain higher-level contexts from a small number of sensor. The sensor data from single node are firstly classified by C4.5 Decision Tree or AdaBoost algorithm, once the confusion matrix of every sensor node have be gotten, the sensor fusion can be performed at the classifier level by calculating the corresponding posterior probability. In order to reduce the level of supervision, this thesis also analyzes the feasibility of active learning for searching most informative samples to be labeled by users in activity recognition. The Experimental results of daily human activity recognition indicate that our algorithm can extract low-level context information from few sensor nodes and then be processed to obtain high-level context information; and the active learning algorithm can detect the most informative unlabeled activity data to ask people to label, so as to learn from large amount of readily available unlabeled data. The research in this thesis owns important practical values in natural human-computer interaction.Finally, it concludes the dissertation briefly, and the future research work is indicated.