The Calculation Of Fast Multi-scale Pyramid Spatio-temporal Intresting Points Based On Prediction Algorithm

In recent years, researchers have done a lot of work on human behavior, in which human behavior often use the body posture to describe transformation relations, and the feature extraction of the human body posture characteristics relies mainly on the video segmentation. The reasons of the person dress, background mutations,illumination change, shadow, object and shelter offender, lead to the attitude of the video analysis divided incomplete, and lost its complete semantic. To solve this problem,we generally use the method of extracting multi-scale pyramid spatiotemporal interesting points to calculate the underlying characteristics,and to make local description for the actions in the video.However, traditional extracting method of multiscale pyramid Spatio-temporal Interesting Points(STIPs) cost larger, takes longer operation, make a poor real-time performance. In order to improve video image multiscale pyramid feature points’ extraction rate, this paper presents a new calculation scheme of fast multi-scale pyramid spatio-temporal interesting points extraction based on a prediction algorithm.To deduce the formula of prediction method and verify its accuracy and timeliness, this paper will do the corresponding detailed calculation and experiment.And there are three experiment databases in this paper which are the INRIA static database,the Weizmann and KTH dynamic database of human behavior.There are three traditional method for the calculation of the Spatio-temporal interesting points,which are Ivan’s method based on Harris corner, Dollar’s method based on Gabor filter on time domain and Willems’ method based on the Hessian matrix. By the experiment of contrasting the three methods, this paper decides to choose Willems’ method to calculate the Spatio-temporal interesting points, and based on which to realize prediction calculation between scales.By the experiments, prove that there is a index coefficient relationship between different scales, and calculate the index coefficient value of DoH feature by the statistical data. With the recursiveness between scales of time and the scale relational expression of two-dimensional DoH characteristics, deduce the prediction expression of Spatio-temporal interesting pointsbetween different scales.By contrasting the prediction algorithm and the traditional calculation method, it is concluded that the proposed prediction has the obvious timeliness. What is more, the accuracy of the prediction algorithm is more than 80%,so it has the very high application value.