针对全卷积孪生网络目标跟踪算法(Siamfc)在严重遮挡、旋转、光照变化、尺度变化等情况下容易出现跟踪失败的问题,提出了一种融合扰动感知模型的孪生神经网络目标跟踪算法。将孪生神经网络提取到的低层结构特征与高层语义特征进行有效融合,以提高特征的表征能力;利用模板自适应策略在线更新模板,以提高算法在遮挡和旋转等情况下跟踪的精确度。与此同时,将基于颜色直方图特征的扰动感知模型引入到算法中,通过加权融合的方式获得目标响应得分图,以此估计出目标的位置,并利用相邻帧尺度自适应策略估计出目标最佳尺度。为验证本文算法的效果,利用公开数据集测试所提算法性能,并与多种跟踪方法进行对比。实验结果表明:在2015目标跟踪标准测试数据集下本文所提算法总体跟踪精确度为0.945,总体成功率为0.929,相比Siamfc算法分别提高了2.9%和2.8%,在无人机航拍测试数据集中本文所提算法也具备较高的精确度与成功率,获得的跟踪效果良好。

Considering that the fully-convolutional siamese network algorithm for object tracking (Siamfc) algorithm is prone to tracking failure in cases such as heavy occlusion, rotation, illumination variation, scale variation, a siamese neural-network object-tracking algorithm with the distractor-aware model is proposed. First, the low-layer structural and high-layer semantic features were extracted from siamese networks; then, they were effectively fused to improve the representation ability of the feature. Second, the template adaptive strategy was used to update the template online to improve tracking accuracy in cases of occlusion and rotation. Simultaneously, the distractor-aware model based on color histogram features was introduced into the algorithm. The target response map was obtained by weighted fusion to estimate the position of the target while the adjacent frame scale adaptive strategy was used to estimate the optimal scale. To verify the effectiveness of the proposed algorithm, its performance was compared with those of various tracking methods on open-source datasets. Experimental results on the standard test dataset of the 2015 ^th object tracking show that the overall tracking accuracy and success rate of the proposed algorithm are 0.945 and 0.929, which is 2.9% and 2.8% higher than those of the Siamfc algorithm, respectively. Further, the proposed algorithm performs with high accuracy and success rate in the aerial test dataset of an unmanned aerial vehicle (UAV).