基于弹载图像的代价敏感与平滑约束结构化SVM目标跟踪方法

Cost-sensitive and smooth-constrained structured SVM object tracking method based on missile-borne images

目标自寻的炮弹打击目标时,传统的目标跟踪方法易受混杂背景的影响,产生模型漂移从而导致跟踪失败。近年来,随着结构化SVM技术的发展,基于结构化SVM的目标跟踪方法能够有效解决复杂背景的问题。弹载成像条件下,在结构化SVM目标跟踪方法的基础上,增加代价敏感权重解决弹载图像中背景混杂所引起的正、负样本不平衡问题,同时利用t时刻超平面法向量w_(t)与t-1时刻超平面法向量w_(t-1)差值的L_(2)范数作为平滑约束抑制模型漂移问题。通过基于对偶坐标下降原理设计了模型的求解算法并实现一种多尺度目标跟踪方法Scale-CS_SSVM。在弹载数据集上进行实验验证,结果表明:与Scale-DLSSVM方法相比,Scale-CS_SSVM在跟踪精度和成功率上分别提高了5.5和5.0个百分点,达到了最优的性能。

When the object self-seeking munition strikes the target,the traditional object tracking method is susceptible to the influence of the mixed background,which produces model drift and thus leads to tracking failure.In recent years,with the development of structured SVM technology,the object tracking method based on structured SVM can effectively solve the problem of complex background.Under the missile-borne environment,the cost-sensitive weights are added to the structured SVM object tracking method to solve the positive-negative sample imbalance problem caused by the background mixing in the missile-borne image,and the L_(2)paradigm of the difference between the hyperplane normal vectors w_(t)at the moment of t and w_(t-1)at the moment of t-1 is utilized as a smoothing constraint to inhibit the model drifting problem.This paper designed a solution algorithm for the model based on the principle of dual coordinate descent and implemented a multi-scale object tracking method Scale-CS_SSVM.Experimental validation on the missile-borne dataset shows that Scale-CS_SSVM achieves the optimal performance with an improvement of 5.5 and 5.0 percentage points in tracking accuracy and success rate,respectively,compared with the Scale-DLSSVM.