Study On Object Recognition And Footstep Detection Of Pattern Recognition

Firstly, this paper discussed object recognition. The Hough transform (HT) is a powerful method for analytic object detection. It is a robust technique for it could get the target through imperfect features. And the algorithm had been implemented to detect lines to find the shoe's direction in the system of shoe sample CAD.The generalized Hough transform (GHT) is not only a method for analytic object detection, but also a powerful method for nonanalytic object detection. Likewise, it has distinct geometric characteristic, and better to noise performance. However, the gradient direction is sensitive to discontinuity and geometric distortion, which degrades the GHT's robustness a lot. To solve the problem, this paper proposed an improved method which characterizing edge points by their normal directions relative to global features. This paper not only presented the reason by principle, but also demonstrated that the improved method do work well for complex objects under severely noise, discontinuity, geometric distortion and overlapping conditions by experimental results.Three methods had been introduced to reduce the 4-D parameter space caused by scale and rotate. Then an approach of building GD-table was proposed to improve one of the three. Furthermore, in order to eliminate false peaks in the Hough domain, the contour sequence stored in model files of the prototype was used to compare with the corresponding sequence of the object. By the determination of the contour error in each region of the Hough domain with high votes, the proposed method was capable of extracting meaningful peaks for further processing.A program for recognizing arbitrary shapes was developed, and had been implemented in the system of shoe sample CAD. Final results had shown that the shoe sample could be detected accurately despite of noise, discontinuity or geometric distortion.Secondly, this paper attempted to study footstep detection. Distinguished from other voices, the footstep has quite an obvious repetitive trait. In this way the cadence of the signal can be used to detect human presence. Five period detection methods had been tested. But none of them had satisfying results. So an improved one level clipper algorithm was proposed basing on footstep's characteristic, which was called twice one level clipper method. It has a good performance in noice-robust. And fifteen different footsteps' periods were detected respectively. Then, the proposed method was used to identify non-footsteps of one hundred different voices, reaching 93% recognition rate.To further identify the non-footsteps with a similar periodicity to footsteps, double decision thresholds endpoint detection, MFCC feature extraction, and DTW recognition method were used. Finally, the recognition rate of the above voices was increased to 98%.The analyses of footstep detection will be valuable into various application fields, such as security, surveillance, and force protection. Also, it can be used to enable video monitoring, which could save medium storage a lot.