多层深度特征的目标跟踪算法研究

Research on object tracking method based on multi-level deep feature

现有的目标跟踪算法大多采用传统的手工特征或神经网络的某一层特征描述目标的外观,不利于跟踪,文中提出一种基于多层深度特征的自适应更新目标跟踪算法。采用经过预训练的深层卷积神经网络分别提取低层和高层信息用以描述目标的空间特征和语义特征,通过对两层特征信息的学习得到两个滤波模板,对应求得两个滤波响应,这两个滤波响应以不同的权重决定最后的跟踪结果。算法中还设计了对目标外观模型和滤波模板的自适应更新方案,能更好地适应目标的外观变化以及遮挡问题。采用多层深度特征描述目标外观,并且利用提取的特征训练两个滤波模板,求滤波响应时采用核相关的方法,增强了跟踪结果的准确性并加快了跟踪的速度。实验结果表明,所提算法与现有跟踪算法相比,可以更好地应对多种挑战因素,跟踪速度也完全能满足实时跟踪任务的需求。

The manual feature or a certain layer feature in neural network is mostly used in existing target tracking algorithms,which is adverse for tracking.Therefore,a multi-layer deep feature based target tracking algorithm with adaptive updating is presented.The pre-trained deep convolutional neural network is used to extract the low-layer and high-layer information respectively to describe the spatial and semantic features of the target.Two filtering templates are obtained by studying the two-layer feature information,and the corresponding filtering responses are acquired.Two filtering responses determine the final tracking result by means of different weights.The adaptive updating scheme of object appearance model and filtering template is designed in the algorithm to adapt to the appearance variation and occlusion of the object.The multi-level deep feature is used to describe the object appearance.The extracted feature is adopted to train two filtering templates.The kernel correlation method used to solve the filtering responses can improve the accuracy of tracking results,and quicken the tracking speed.The experimental results show that,in comparison with the existing tracking algorithms,the proposed algorithm can deal with the multiple challenges perfectly,and its tracking speed can fully meet the requirement of real-time tracking task.