Gait Recognition Based On Haptic Force Information

In the information age, safety certification requirements have become increasingly popular and stringent. Traditional authentication technologies such as ID card, tailor-made key and password can't meet the needs of new situation, because they can easily be forged. Biometric recognition technologies are pinned high expectations. Fingerprint, palmprint, hand geometry, iris, face, DNA, voice and signature, etc. play an important role in many areas, as ensuring personal information security, preventing terrorist incidents and combating economic crimes and so on. However, these biometric recognition technologies belong to short-distance human identification technologies, the targeted subjects are required to closely cooperate with the measurement scheme, which can easily lead to human rights disputes. Gait recognition is a distinctive human identification technology at a distance, it has advantages as non-invasive detection and concealed installation, so it won't lead to human rights disputes and the targeted subjects won't disguise.The research results of cognitive science, sports biomechanics and criminal investigation show that haptic information of gait contains abundant walking patterns and habits, which have potential of identifying human. In this thesis, we systematically studied human identification based on haptic force information of gait from the perspective of sports biomechanics, we studied the methods of ground reaction force measurement, data preprocessing, gait feature extraction, gait feature selection and classification. Human identification based on haptic force information of gait has another distinctive advantage, it is not affected by complex background and shelter, it will become an important research direction in the field of human identification at a distance, this research has important theoretical significance and broad application prospect. The main works and innovative achievements in this thesis are introduced as follows.(1) We analyzed the feasibility and the research significance of human identification based on haptic force information from multiple perspectives, and made the research scheme. We built a haptic force information collection walkway using five self-developed force platforms, which are alternately embedded under ground in equal interval style. This walkway can acquire continuous intact ground reaction force. We built three gait databases (ITCSH Gaitâ… , ITCSH Gaitâ…¡and ITCSH Gaitâ…¢), which has different amounts of samples and different information contents. In addition, ITCSH Gaitâ…¢has synchronous gait image sequence from five view-angle. The three gait databases filled the international gaps in this field, and will play a positive role in promoting the gait recognition research.(2) To ensure data quality, we chose wavelet thresholding for GRF denoising. Waveform alignment method was proposed for enhancing the comparability of frequency-domain gait features extracted by wavelet packet decomposition. Re-sampling method was proposed for effectively expanding gait samples. Experimental results proved that the three data preprocessing methods are beneficial to improve the classification performance. We firstly demonstrated by experiments that using anyone data normalization method of min-max normalization, z-score normalization and weight normalization for reducing magnitude difference of GRF components will reduce classification performance in varying degrees, then we analyzed the possible reasons.(3) A time-domain gait feature extraction method based on waveform characteristic points detecting was proposed, this method can extract temporal-spatial parameters and kinetic parameters from GRF, these parameters can reflect the holistic and detailed gait characteristics, and we built some derivative kinetic parameters by computational method. All of the parameters constructed the time-domain gait features which have clear physical meanings, then they are used to identify human. We suggested use ICC-based reliability verification method and Kruskal-Wallis rank sum test method to validate the stability and uniqueness of the time-domain gait features, and experimental results proved that the time-domain gait features are stable and unique. Wavelet packet decomposition was utilized to extract frequency-domain gait features, contrast was found that time-domain gait feature extraction method based on waveform characteristic points detecting can help to understand the principle of gait recognition.(4) Correlation-based feature selection (CFS) method that combining with correlation measure and best first search strategy was proposed, this method can reduce redundant gait features of class independent by the greatest extent. SVM-Wrapper feature selection method that utilizing classification performance feedback was used. The two feature selection methods can really reduce the number of features required extraction. On the basis of theoretical analysis and experiments, two combinatorial feature selection methods called'CFS+PCA'and 'CFS+SVM-Wrapper'were proposed. Support vector machine based on Gaussian radial basis function (RBF-SVM) was utilized to test the performance of each feature selection method listed above. Experimental results proved that CFS and SVM-Wrapper has good performance,'CFS+PCA'and'CFS+SVM-Wrapper'can reduce feature dimension by the greatest extent and maintain good classification performance. Meanwhile, classification results demonstrated that RBF-SVM can well classify and recognize the gait samples even though in small training sample size condition, its performance was better than KNN.(5) It was founded that false recognition mainly occurred when subjects have similar weight and walking speed varies greatly. So, we suggested utilize support vector regression (SVR) model to real-time regress the subject's weight and height from time-domain gait features, it can provide prior knowledge for improving the gait recognition performance and forensic applications.(6) Contrast experiments showed that the classification performance of RBF-SVM classifier decreased significantly when walking speed varied greatly. We proposed distributed voting classification method and firstly utilized rotation forest (a multiple classifier system) to deal with the gait recognition challenge while walking speed changes. Experimental results proved that rotation forest can well recognition people when walking speed changes less, distributed voting classification method can have relatively high recognition performance while walking speed changes more.