自适应尺度特征融合的异常重检跟踪算法

Self-adaptive scale feature fusion of anomaly re-detection tracking algorithm

针对核相关滤波(KCF)跟踪算法中无法处理目标尺度变化以及单一特征的局限性问题,并避免在目标遮挡、形变等情况下跟踪器对目标的丢失,提出一种自适应尺度特征融合的异常重检(SFAR)跟踪算法。通过融合多尺度及颜色特征后的核相关滤波器,检测跟踪目标并得到多尺度图像,然后根据提取的多尺度图像去训练成多尺度目标模型,在其过程中引入峰值的异常值判定来矫正,若发生异常峰值则判定目标跟踪丢失,后进入重检机制。在公开的OTB-50数据集上测试,实验结果表明,改进算法精度为0.740,成功率为0.591,跟踪效果得到显著提升。

Aiming at the problem of the inability to deal with the scale variation and the limitations of a single feature in kernel correlation filter(KCF),simultaneously,to avoid the target loss due to occlusion,scale variation and so forth,an anomaly rechecking tracking algorithm based on adaptive scale feature fusion was proposed.By combining the multi-scale and color features of the kernel correlation filter,the tracking target was detected and multi-scale images were obtained,then according to the extracted multi-scale image,the multi-scale model was trained,in this process the peak outlier determination was introduced to correct target location,if the abnormal peak value was judged,the target tracking was lost and then the re-detection mechanism was taken.The proposed method was tested on the open OTB-50 dataset,its results show the improved algorithm’s accuracy is 0.740,and the success rate is 0.591,the tracking effect is improved significantly.