基于分层粒子滤波的地标检测与跟踪

Landmark detection and tracking based on layered particle filter

针对直升机运动平台获得的地标图像序列高噪声且帧间位移大等难点,提出一种改进的局部二值模式向量,利用集成学习算法训练分类器区分地标模式和背景模式.采用分层粒子滤波算法有效地结合纹理和颜色2种特征,将地标检测与跟踪有机地融为一体.该算法首先选择似然值均值作为观测值,并利用均值漂移算法将粒子移到高似然值区域,然后选择颜色直方图作为观测,逐步将粒子移向高权重区域,最后通过聚类算法估计地标数目与位置.实验验证了该算法的有效性和鲁棒性.

To overcome high noise and large inter-frame displacement in landmark image sequences captured from a moving-helicopter platform,an improved local binary pattern descriptor was proposed to discriminate patterns of landmark and those of background using the classifier trained by the ensemble method.The layered particle filter is adopted to fuse the color-based and texture-based features,and the detection and tracking are combined in a principled way.Likelihood mean is chosen as observations first,and particles are moved towards their local modes by mean shift in likelihood image.Color histogram is then used to sample particles with high weight gradually.Finally clustering algorithm is applied to estimate the number and positions of landmarks.Experiment on real image sequences demonstrated that the proposed algorithm is effective and robust.