基于定位-分类-匹配模型的目标跟踪方法

Target Tracking Method Based on Location-Classification-Matching Model

近年来卷积神经网络框架被成功地应用到目标跟踪领域,并取得了较为稳健的跟踪结果。基于此思想,提出一种基于定位-分类-匹配模型的目标跟踪方法。首先,在定位模型中,利用前一帧的位置信息预测当前帧中的候选目标区域。然后,采用已训练的深度特征对候选区域进行类间筛选,选出N个次优目标区域。最后,利用常规颜色特征对次优目标区域进行类内寻优匹配,从而确定最终的跟踪目标。与此同时,分别对定位、分类中的网络进行更新,并对建立的匹配模型进行在线实时更新,使得其对目标的描述更加准确。在OTB50和OTB100标准数据库上进行实验测试,结果表明,提出的跟踪方法在快速运动、相似物体干扰、复杂背景等条件下具有较好的跟踪稳健性。

Recently, the framework of convolution neural network has been successfully applied to the target tracking, and has achieved robust tracking results. On the basis of this conception, a target tracking method based on location-classification-matching model is proposed. First of all, in the location model,the candidate target region of the current frame is predicted by using location information of previous frame. Secondly, the trained deepth features are used to inter-class screen the candidate regions, and N sub-optimal target regions are selected. Finally, we use conventional color features to perform intra-class optimization matching for sub-optimal target regions, so as to determine the final tracking target. Meanwhile, the network in the location and the classification is updated separately, and the established target model is updated online and real-time to ensure that the model describes the target accurately. Experimental tests are performed on OTB50 and OTB100 standard databases, the experimental results show that the proposed tracking method has better tracking robustness under the conditions of fast motion, similar object interference, and complex background.