自适应轮廓的变分水平集复杂背景多目标检测

Adaptive Contour Based Variational Level Set Model for Multiple Target Detection in Complex Background

无需重新初始化的变分水平集模型能够避免经典水平集模型的重复初始化步骤,进而简化计算,缩短检测所需时间,同时能够有效利用图像的边缘梯度信息,从而准确定位图像的局部结构.但该模型不能自适应地获得初始化曲线,水平集的拓扑结构也无法改变,不能解决多个目标的检测问题.针对以上问题,提出了一种基于自适应轮廓的变分水平集复杂背景多目标检测方法,该方法采用帧间差分算法与K-means聚类算法相结合,以获得多个运动目标的初始化曲线,通过形态学方法来降低图像噪声的干扰,从而快速自适应地估计复杂背景下运动目标的位置和轮廓大小.该算法进一步对无需初始化的变分水平集进行改进,将其由单目标检测模型扩展为多目标检测模型,并修正原模型难以处理图像灰度不均匀的问题,最终实现对复杂背景下多个目标的检测.在标准数据库和实际数据集上的测试结果表明,所提方法能够准确地定位不同尺度和灰度目标的轮廓,从而提高算法的演化迭代效率及准确性.

Comparing with the classical level set, the variational level set without re-initialization can avoid repeating initialization, which greatly reduces the algorithm＇s running time while using the edge gradient information of images to accurately capture the local structure. However, this model cannot adaptively obtain initial curve, and the model＇s topology cannot be changed to detect multiple targets. To solve the problems above, this paper proposes an adaptive contour based variational level set model for multiple target detection in complex background. First, the inter-frame difference algorithm is combined with K-means clustering algorithm to obtain multiple initialization curves, and then the noise is reduced by morphology method. This can estimate the position and the size of the moving target in complex background. The variational level set without re-initialization is further extended to multiple targets from single target, and the model＇s ability is improved to deal with the images of non-uniform gray. Experiments on standard database and real scene data sets indicate that the proposed method can accurately locate targets contours of different scales and gray to improve the evolution efficiency and accuracy of the algorithm.