摘要
通过移动机器人的视觉系统可以实现动态目标的检测与跟踪。提出一种基于改进的高斯混合模型(GMM)的实时动态目标检测算法,算法引入分块思想,在模型更新过程中动态调整GMM分布数目和学习率,通过改进匹配准则来减小误检率(FPR)和漏检率(FNR)。在目标检测的基础上,采用一种融合均值偏移(MS:Mean Shift)和粒子滤波(FP:Particle Filter)的算法对目标实时跟踪,在利用MS算法获得的最优候选区域周围散布采样粒子,根据偏移向量的大小自适应调节粒子数目,不仅具有较快的收敛速度,且对遮挡具有较好的鲁棒性。实验结果表明,将两种改进后的算法应用于移动机器人的视觉系统中,能够对动态场景中的动态目标实时检测与跟踪,较传统算法在实时性和精确性上均获得一定提高。
Mobile robot can detect and track moving object through its vision sys tem.In order to reduce both false positive rate （FPR） and false negative rate （FNR） of detection,an improved Gaussian mixture model （GMM） for dynamic target real-time detection is presented in this pape r.In this improved algorithm,the concept of block is introduced,then the number and learning rate of Gaussian mixture model can be adjusted dynamically during the updating proces s,and finally the matching criterion of GMM is improved to decrease the false positive rate.After detection,a fusion tr acking algorithm based on mean shift and particle filter （PF） is adopted to promote real-time capability of tr acking system,the particles are scattered around the optimal candidate region obtained by mean shift algorithm, and the quantity of particles is adaptively changed according to mean shift value.This fusion algor ithm combines the advantages of the mean shift and particle filter,has a fast converg ence speed and is robust to occlusion.Experiments implemented on video frames show that the two propose d improved methods not only give higher accuracy than traditional algorithms,but also can detect a nd track the moving object efficiently in mobile robots′ vision system.
出处
《光电子．激光》
EI
CAS
CSCD
北大核心
2013年第11期2198-2204,共7页
Journal of Optoelectronics·Laser
基金
国家自然科学基金(61001174)
天津市自然科学基金(13JCYBJC17700)资助项目
