介绍了一种采用小波变换抑制港口动态背景的方法。在此基础上,运用基于Chan-Vese模型的水平集方法实现图像区域分割和粒子区域的特征提取,提取的特征包括结构和灰度特征,结合这些粒子区域特征和先验知识,在对标准粒子滤波进行改进的基础之上,实现多目标跟踪中的数据关联和目标聚类。同时,在多目标跟踪中会出现"新目标出现、目标消失、多目标合并与分裂、目标受遮挡"等现象,根据图像能量和轮廓形状间距离的变化来判断和解决上述现象带来的误跟踪问题。通过对实际港口背景条件下的红外序列图像进行多目标跟踪实验,验证了所提方法的可行性和有效性。
An efficient approach based on wavelet transform was presented to restain the harbor dynamic background.And then,image region segmentation and region features extraction were realized by level set method based on Chan-Vese model.Region features included shape structural information and gray level.Combining those features with dynamic priors,data association and those targets cluster were realized by using the improved particle filter.In addition,"new target emerging,target disappearing,multiple targets merging and splitting as well as target occultation" phenomena occasionally arose during multi-target tracking.The false tracking problems caused by the above phenomena were judged and solved by the variation degree of image energy and contour distance.The proposed approach was validated to track multi-target effectively by using actual infrared image sequences against complex harbor background.Experimental results indicate the feasibility and effectiveness of the proposed method.