基于残差网络运动场景分类的目标跟踪算法

Target tracking algorithm based on residual network movement scene classification

在目标跟踪中,目标本身容易发生变化,且目标的运动场景是复杂多样的,而不同网络训练的跟踪模型在跟踪同一场景的目标性能会有较大差异,进而使很多算法的跟踪效果不太理想。针对这一问题,文章提出了一种基于目标运动场景分类的目标跟踪方法,解决了单一模型在应对目标处于不同复杂场景中性能不稳定的问题。该算法利用残差网络对目标运动场景进行分类,并且使用迁移学习提高了分类的准确率,之后选取合适的网络模型对目标进行跟踪。在UAV123数据集上与原始的单一模型进行对比实验的跟踪结果表明,改进的算法能够有效地提高目标跟踪的成功率和精度,在OTB100数据集上与其他跟踪器进行比较,跟踪效果均优于其他跟踪器。

In target tracking,the target is prone to change and the movement scene of the target is complex and diverse.Tracking models which trained by different networks have large differences in targets tracking performance in same scenes,which makes the tracking effects of many algorithms are unsatisfactory.Aim at this problem,this paper proposed a target tracking method based on target movement scene classification,which solves the problem of unstable performance of a single model in dealing with those targets in different complex scenes.The algorithm uses the residual network to classify the target movement scene,at the same time,raise the precision of classification by transfer learning,and then selects a suitable network model to track the target.The tracking results of the comparison experiment with the original single model on the UAV123 data sets show that the improved algorithm can effectively raise the success rate and precision of target tracking.Compared with other trackers on the OTB100 data set,the tracking effect is better than other trackers.