Hand Detection Based On Enhancement Of Sampling Probability Density

Object detection and recognition technologies play a key role in human-computerinteraction and other relative fields. The precision and computation cost duringdetection and recognition are two significant criterions in human-computer interactionsystem. Because of the in-plane rotation of target objects, objects in any certain anglerange are considered as one new class. The variation of angle requires the system todeal with many single-angle object detections. This causes the great increase ofcomputation cost while decreasing the whole quality of system and user experience.This paper focuses on the problem of hand detection with multiple angles in plane.The uncertainty of angle of hands requires the system to detect many hands with anysingular angle, which brings more computation cost in multiple angle hand detectioncomparing with singular angle hand detection. Sliding window strategy andtraditional particle window strategy are introduced in this paper firstly, with theanalysis of the two strategies sequenced. A new method is proposed in this paper tosolve the problem. Based on the particle window strategy, the new method takesadvantage of correlative information among the sampling probability density functionof different angular classifiers. Using enhancing the sampling probability density canmake the process of iteration in particle window strategy faster, which decreases thecomputation cost with same precision.The main contributions of this paper are as follows:(1) Extend the concept ofregion of support from the dimension of space and scale to angle and prove theexistence of relative information among different classifiers.(2) In order to decreasethe computation cost, this paper presents a new method which boosts the samplingprobability density function of relative classifiers to impose the relative information inthe process of detection.(3) The new method of hand detection in multiple anglesprovides a new idea to the work of detection which includes in-plane rotation orrelative classifiers.