摘要： 跟踪-学习-检测(TLD)目标跟踪算法能够实现长时间的在线目标跟踪,但当目标平面旋转发生形变以及目标被严重遮挡时,TLD算法在跟踪过程中会产生跟踪漂移。针对上述问题,在TLD算法的跟踪模块上使用稀疏原型进行跟踪,提出一种稀疏原型(SP)-TLD目标跟踪算法。当出现由于平面旋转引起的目标形变时,通过仿射变换变化坐标位置,能够准确跟踪目标避免产生跟踪漂移。在目标被严重遮挡时,根据目标的主成分分析基向量和琐碎模板判断目标未被遮挡及被遮挡部分,从而识别出被遮挡的目标。实验结果表明,与TLD算法相比,SP-TLD算法具有更高跟踪准确率和更强鲁棒性。
关键词:
目标跟踪,
稀疏原型,
仿射变换,
主成分分析基向量,
琐碎模板
Abstract: Tracking-learning-detecting(TLD) object tracking algorithm can achieve a long time online tracking.But when the appearance changes of the object by the plane rotation and the object is occluded,the TLD tracking algorithm tracks drift in the process of tracking.Aiming at above these problems,the tracking module of the TLD algorithm adopts sparse prototype,and the Spares Prototype(SP)-TLD object tracking algorithm is proposed.When the object deformation caused by the plane rotation is present,the SP-TLD object tracking algorithm accurately tracks the object by changing the coordinate position of the affine transformation and avoids tracking drift.When the object is occluded,the algorithm determines the unoccluded part and occluded part according to the Principal Component Analysis(PCA) basis vectors and trivial templates of the object and tracks the object accurately.Experimental results show that the SP-TLD algorithm has higher tracking accuracy and stronger robustness that the TLD algorthm.
Key words:
object tracking,
Sparse Prototype(SP),
affine transformation,
Principal Component Analysis(PCA) basis vector,
trivial template
中图分类号: