复杂场景下自适应特征融合的图像运动目标跟踪算法研究

Research on Image Moving Target Tracking Algorithm Based on Adaptive Feature Fusion in Complex Scenes

针对目标跟踪所面临的尺度变化、快速运动导致的跟踪漂移或失败问题,提出一种复杂场景下自适应特征融合的图像运动目标跟踪算法。本文分别设计了目标分类和目标估计模块,并将其有效结合。在目标分类模块,设计了一种自适应特征融合机制。该机制融合了多层深度特征以实现有效的在线跟踪。此外,设计的联合更新策略通过优化投影矩阵层和相关层,在处理运动模糊、严重目标形变时具有更强的鲁棒性。在目标估计模块,引入IoU(Intersection over Union)最大化的理念,预测目标和估计边界框之间的IoU分数,在跟踪过程中,通过使用梯度上升最大化IoU分数来估计目标状态,获得更加精确的边界框。实验结果表明,本文所提出的算法具有更出色的跟踪性能,其在OTB100,UAV123及LaSOT数据集上的S AUC分别为70.1%,47.6%和51.6%,优于其他相关算法。

Aiming at the problems of tracking drift or failure in target tracking for the scale change and fast motion,a image moving target tracking algorithm based on adaptive feature fusion in complex scenes is proposed.In this paper,the target classification module and target estimation module is designed respectively and combined effectively.In the target classification module,an adaptive feature fusion mechanism is designed,and it integrates multi-layer depth features so as to achieve effective online tracking.Moreover,the designed joint update strategy is more robust in dealing with motion blur and severe target deformation by optimizing the projection matrix layer and the correlation la-yer.In the target estimation module,the concept of IoU(Intersection over Union)maximization is introduced to predict the IoU score between bounding boxes and the estimation target.During the tracking process,the target state is estimated by using gradient ascent to maximize the IoU score to obtain a more accurate bounding box.Experimental results show that the proposed algorithm has excellent performance,with S AUC of 70.1%,47.6%,51.6%on OTB100,UAV123 and LaSOT datasets,which is superior to other related algorithms.