Research On Footprint Recognition Algorithm Based On Local Spectrum Feature

With the rapid development of identity recognition technology, demand for this kind of technology is increasing and the requirement of its reliability is higher. Many authentication technologies based on biometric have gotten rapid rise, among which behavior characteristics, as special biological characteristics, due to their security, concealment and flexibility, have more and more attention by scholars both at home and abroad. Footsteps, as one of the main behavior characteristics, can be used to achieve the aim of identifying different identities. Information in many aspects such as character, age and gender and so on is contained in the footsteps, so accurate identification of footsteps will be of great value and broad application prospects in security service and supervision as well as individual behavior interpretation.For a long time, footstep identify research is mainly focused on the study of acoustic characteristics, which can be very sensitive to environmental noise interference, so the environmental noise is very strictly limited,(that is, in a quiet indoor environment). This is unrealistic in practice, even in indoor environment, other sounds are always mixed in the footsteps. Therefore, how to extract footstep features which can accurately reflect the characteristics of human walking, and have strong robustness to noisy environment remains the main problem in this area. According to the relationship between footstep spectrogram and time-domain footstep signal, a footstep recognition algorithm based on local spectrogram features is put forward in this paper. Main work and progress are as follows:(1) Introduced the equipment and environment of the recording of footsteps, and the acquisition of footstep database as well as the generation of training and testing samples. Furthermore, how to generate a spectrogram from footstep signal is also expounded.(2) Aimed at the problem that traditional speech feature extraction methods based on acoustic parameters are of the poor robustness to the noise, a new footstep recognition algorithm is proposed based on the locality of footstep distribution in time and frequency compared with background sound and combining with the visualization of footsteps spectrograms. This method can be divided into three steps, first of all, in order to enhance robustness to white noise and Gaussian noise, spectral subtraction is used in pretreatment of spectrogram. Then, the process can be divided into two parts, training and identification. In the process of training, calculate the logarithm energy of short-time Fourier spectrum of each training sample acquainted in a quiet environment. The method proposed in digital image of key points detection and characterization is used here for key points detection in the two-dimensional spectra, and so form characterization of every key point, namely the local spectrum features of the key point.(3) In identification process, key points are detected in logarithm energy short-time Fourier spectrum of unknown sample containing noise, forming feature representations of key points. Considering that footsteps, as a kind of behavior characteristic, do not have good stability. Even the local spectrum features of footsteps from the same person in different times will change to some extent. On the other hand, the similar frequency component in background will cause interference to local spectrum features in the same time frame. Thus the minimum error Bayes decision theory can be used classify these feature representations, category mark as a result of unknown sample classification.(4) The proposed algorithm is compared with traditional footstep recognition algorithms based on the acoustic characteristics, and experiments are simulated respectively in a pure footstep environment and various kinds of noisy environment. Experimental results show that the recognition accuracy of the algorithm proposed in this paper is significantly higher than that of the existing algorithm in different kinds of noise or different SNR environments and the robustness to the different background noise and environmental noise is significantly better than the existing algorithms.