基于先验显著性的卫星视频微小运动车辆检测

Detection of tiny-moving vehicles in satellite video based on priori-significance

卫星视频中微小运动车辆由于特征信息少,使得当前基于特征的方法很难获得良好结果,为了应对卫星图像中复杂的图像环境以及目标微小的像素占比,提出了一种基于多种图像先验信息约束显著性的运动车辆检测方法。首先利用简单线性迭代聚类算法对视频帧生成数量适宜结构紧凑的超像素,并基于人眼显著性模型计算各超像素的边界相关性;同时采用背景建模生成图像的运动热图并完成基于超像素的运动区域估计;结合这两种先验约束来最优化检测结果,从而获得较好的目标检测效果。最后,以吉林一号卫星视频中全球不同城市道路场景为实验数据,对比了当前多种运动目标检测算法,实验证明,提出的算法能在检测准确率和召回率达到85%的情况下保持低于10%以内的虚警率,并且在卫星角度偏移的图像环境也有一定的抗干扰能力。

Owing to the lack of feature information in satellite videos,obtaining good results using the cur⁃rent feature-based method is difficult.Therefore,in this study,a target detection method based on the im⁃age prior information constraint significance is proposed.First,a simple linear iterative clustering algo⁃rithm is used to generate a suitable number of superpixels,with compact structures for video frames,and the boundary correlation of each superpixel is calculated based on the human eye significance model.Simul⁃taneously,background modeling is used to generate a motion heat map of the image.Then,motion region estimation is derived based on superpixels.Finally,the two prior constraints are combined to optimize the detection results to obtain a better target detection effect.Different urban road scenes from the“jilin-1”sat⁃ellite video are taken as examples,and several existing algorithms are compared.Experiments prove that the proposed algorithm can achieve a false alarm rate of less than 10%when the detection accuracy and recall rate reach 85%.The algorithm also has good anti-interference ability in an image environment when the angle deviation of the satellite is considered.