基于卡尔曼滤波的SiamRPN目标跟踪方法

SiamRP network for object tracking based on Kalman filter

基于深度学习的视觉跟踪方法在多个基准数据库上取得了很好的跟踪性能,特别是基于Siamese框架的目标跟踪方法取得了突破性的进展。为了提高跟踪效果,有效解决跟踪过程中干扰和遮挡问题,本文提出了一种基于卡尔曼滤波的SiamRPN(Siamese+RPN)目标跟踪方法。首先,利用训练好的SiamRPN跟踪算法和卡尔曼滤波跟踪模型分别对目标物体进行跟踪,得到2种跟踪算法跟踪结果的置信度,然后,基于置信度加权融合模型得到最后的跟踪框。卡尔曼滤波器可预测目标在一定遮挡干扰等情况下的位置,SiamRPN算法利用区域候选网络RPN将每一帧的跟踪转换为一次局部检测任务,快速准确地得到跟踪框的位置和尺度,提出的算法避免了使用常规的低效费时的多尺度自适应方法,融合了2种优秀跟踪算法的优点,不仅跟踪速度较快,而且抗干扰和遮挡能力明显提高。在经典数据库上的实验验证了提出的算法明显提高了目标运动较快、干扰较强和有遮挡情况下的跟踪效果,在速度没有明显下降的前提下,成功率和精度等多个性能指标均有较大的提升。

The visual learning method based on deep learning has achieved good tracking performance on multiple benchmark databases.Especially,the object tracking method based on Siamese framework is a breakthrough.In order to improve the tracking effect and solve effectively the interference and occlusion problems in tracking,a SiamRPN(Siamese+RPN)object tracking method based on Kalman filtering is proposed.Firstly,the trained SiamRPN tracking algorithm and the Kalman filter tracking model are used to track the object respectively,and the confidence of the results of the two tracking algorithms is obtained.Then,the final tracking frame is obtained based on the confidence weighted fusion model.The Kalman filter can predict the position of the object under certain occlusion interference.The regional candidate network RPN in SiamRPN algorithm is used to convert the tracking of each frame into a local detection task,and obtain the position and scale of the tracking frame both quickly and accurately.The conventional inefficient and time-consuming multi-scale test and online fine-tuning is abandoned.The new method includes the advantages of two excellent tracking algorithms.Thus,not only the tracking speed is fast,but also the anti-interference and occlusion capabilities are significantly improved.Experimental results on the classical database verify that the proposed algorithm significantly improves the tracking effect in the condition of fast object motion,strong interference and occlusion.The performances such as success rate and accuracy are greatly achieved without decreasing the tracking speed obviously.