超像素特征的运动目标检测算法
Moving object detection algorithm with superpixel features

在像素级的背景建模方法中, 由于其反映的只是时间上的连续性, 没有考虑到空间上的相关性, 所以会导 致检测目标不完整, 或检测目标呈碎片化的结果, 不利于后续的识别或跟踪. 为此, 本文首先针对ViBe算法对于动 态背景不鲁棒的问题进行了改进, 利用样本集的标准差作为动态背景度量值, 实时更新距离阈值和背景模型更新 率, 达到对动态背景的鲁棒性; 同时引入了超像素特征, 提出了基于超像素特征的运动目标检测算法. 由于超像素 分割具有较好的边缘信息同时超像素数目可控, 所以根据SLIC0超像素分割算法提取超像素特征, 将超像素块中的 像素均值作为超像素特征值, 并引入到改进的ViBe算法框架中; 由于超像素分割的数目并不是固定不变的, 所以本 文使用初始种子点位置的超像素特征构建背景模型并进行运动目标的检测. 实验表明, 该方法检测结果具有良好 的目标边缘信息并可以有效抑制动态背景的干扰.
The pixel-level background modeling algorithms take into account only the time continuity, without considering spatial correlation. So it causes incomplete detection, which is not conducive to subsequent identification or tracking task. To deal with the above problems, we do the following work. Firstly, the ViBe algorithm is not robust to dynamic background, so a method to measure the dynamic background complexity is proposed by using the samples’ standard deviation. And the value is used to update online the background modeling update rate and distance threshold, which can achieve robust detection. Secondly, amoving object detection algorithm based on superpixel features is proposed. The superpixel segmentation results have many advantages, such as better edge information and controllable superpixel number. And superpixel features are extracted using SLIC0 segmentation algorithm, the pixels’ mean value in superpixel blocks is used as the feature value, and the superpixel features are used in pixel-level background modeling framework. Finally, due to the number of superpixel features is unstable, so the superpixel features, which are used to construct the background modeling, are located in the same position of initial seed points. Experiments show that the proposed method can obtain better objects edge information and it can effectively reduce the effects of dynamic background.