基于AdaBoost检测器的似然估计方法

Likelihood estimation based on AdaBoost detector

本文提出利用Gentle AdaBoost（GAB）训练一个层叠结构的目标检测器,然后基于训练出的检测器结构引入两种策略,设计了5种应用于后续粒子滤波跟踪的似然函数.为估计目标出现的概率,提出了两种构造似然函数的策略：层内概率统计（PIS）策略和层间概率统计（POS）策略.PIS表示在同一层内每个弱分类器的实数输出的概率统计,POS为实现层叠检测器在检测时所到达深度的概率统计.基于这两种策略,设计出了5种似然函数的形式：基于层叠结构的层内概率密度估计似然函数（PIS-CA）、基于合成结构的层内概率密度估计似然函数（PIS-EA）、层间概率密度估计似然函数（POS）、顺序组合层叠检测器的层内概率密度估计似然函数（S-PIS-POS）和逆序组合层叠检测器的似然函数（A-PIS-POS）.实验表明,所定义的似然函数可以很好地估计目标出现的概率,在目标出现的区域比背景区域具有更大的置信度,整合PIS和POS两种策略的似然函数具备最优的性能.

This paper presents a novel likelihood estimation which can be used for particle filter based object tracking. The likelihood estimation is built upon the cascade object detector trained with Gentle AdaBoost （GAB） to capture the probability of the existence of object. Two strategies are adopted to construct the likelihood functions： Probability-Intra-Stage （P IS） corresponding to real＇output of each weak classifier in the same stage, and Probability-Outer-Stage （POS） corresponding to the depth reached in the cascade detector. Five kinds of likelihood functions, PIS-CA, PIS-EA, POS, SPIS-POS, and A-PIS-POS, are thus proposed based on the trained GAB detector. The experiment shows that the likelihood functions can probabilistically characterize the existence of object, with much higher confidence value in object regions than that in the background, and that the integral strategy of PIS and POS is the best choice.