融合SVM的多特征DSST目标跟踪算法

Multi-feature DSST Target Tracking Algorithm Based on SVM Fusion

为解决DSST算法多尺度搜索策略跟踪时目标出现严重遮挡、非刚性形变、目标脱离视场导致的目标外观变化的问题,提出一种将支持向量机(support vector machine,SVM)目标重检测模块融合的算法。提取目标的多种特征然后将这些特征矢量融合以增强目标的特征表达。在DSST算法的位置和尺度滤波器的基础上,新增目标外观滤波器,利用训练好的SVM全局搜索目标。采用不同大小的窗口采样来训练相关模型并建立一个SVM的最优分类面,通过SVM对丢失后的目标进行重检测。实验结果表明,改进算法比DSST算法在对目标受到遮挡、目标非刚性形变等问题上的鲁棒性能均有提高。

In order to solve the problems of object appearance change caused by severe occlusion,non-rigid deformation and object departure from the field of view in multi-scale search strategy tracking of DSST algorithm,an algorithm was proposed to fuse the support vector machine(SVM)object re-detection module.Multiple features of the target are extracted and these feature vectors are fused to enhance the feature expression of the target.Based on the position filter and scale filter of DSST algorithm,the target appearance filter is added,the trained SVM is used to search for the target globally.Different window samples are used to train relevant models and establish an optimal classification surface of SVM.The missing target is re-detected by SVM classifier.The experimental results show that the improved algorithm has better robust performance than DSST algorithm on such problems as target occlusion and non-rigid deformation.