Auto-attentional mechanism in multi-domain convolutional neural networks for improving object tracking

Purpose-Multi-domain convolutional neural network(MDCNN)model has been widely used in object recognition and tracking in the field of computer vision.However,if the objects to be tracked move rapid or the appearances of moving objects vary dramatically,the conventional MDCNN model will suffer from the model drift problem.To solve such problem in tracking rapid objects under limiting environment for MDCNN model,this paper proposed an auto-attentional mechanism-based MDCNN(AA-MDCNN)model for the rapid moving and changing objects tracking under limiting environment.Design/methodology/approach-First,to distinguish the foreground object between background and other similar objects,the auto-attentional mechanism is used to selectively aggregate the weighted summation of all feature maps to make the similar features related to each other.Then,the bidirectional gated recurrent unit(Bi-GRU)architecture is used to integrate all the feature maps to selectively emphasize the importance of the correlated feature maps.Finally,the final feature map is obtained by fusion the above two feature maps for object tracking.In addition,a composite loss function is constructed to solve the similar but different attribute sequences tracking using conventional MDCNN model.Findings-In order to validate the effectiveness and feasibility of the proposed AA-MDCNN model,this paper used ImageNet-Vid dataset to train the object tracking model,and the OTB-50 dataset is used to validate the AA-MDCNN tracking model.Experimental results have shown that the augmentation of auto-attentional mechanism will improve the accuracy rate 2.75%and success rate 2.41%,respectively.In addition,the authors also selected six complex tracking scenarios in OTB-50 dataset;over eleven attributes have been validated that the proposed AA-MDCNN model outperformed than the comparative models over nine attributes.In addition,except for the scenario of multi-objects moving with each other,the proposed AA-MDCNN model solved the majority rapid moving objects tracking scenarios and outperformed than the comparative models on such complex scenarios.Originality/value-This paper introduced the auto-attentional mechanism into MDCNN model and adopted Bi-GRU architecture to extract key features.By using the proposed AA-MDCNN model,rapid object tracking under complex background,motion blur and occlusion objects has better effect,and such model is expected to be further applied to the rapid object tracking in the real world.

基于改进MDNet的视频目标跟踪算法

目标跟踪算法面对的突出问题之一是正负样本不均衡,正样本极度相似.针对跟踪更新过程中正样本不丰富的问题,本文基于多域卷积神经网络(MDNet)算法,提出了一种改进MDNet的视频目标跟踪算法,首先改进原算法中候选框的选取策略,提出了一种基于候选框置信度与坐标方差阈值判断相结合的模型更新方法,其次将原算法的交叉熵损失函数改进为效果更好的focal loss损失函数.实验结果表明在相同实验环境下本文算法相较于MDNet算法在跟踪准确率和成功率上均有明显提高.

基于FasterMDNet的视频目标跟踪算法

多域卷积神经网络(MDNet)算法在卷积层采用选择性搜索的方式来提取候选框,因此它没有共享完整图像特征,从而导致在线视频目标跟踪速度慢。针对这个问题,提出一种快速多域卷积神经网络(FasterMDNet)视频目标跟踪算法。FasterMDNet是建立在MDNet基础上的一种模型,在卷积层后面引入RPN(Region Proposal Network)网络,优化了损失函数,共享完整图像卷积特征,加快候选区域建议框(ROI)更高效的生成;为了更好地获得目标和背景信息表示,在RPN网络后加入ROIAlign层,对提取的候选区域建议框特征图用双线性插值方法来提高感受野的分辨率。该算法对目标跟踪基准数据集OTB2013、OTB2015、VOT2016进行了评估,并与前沿的跟踪算法做对比,实验结果证明,该算法跟踪准确率优于其他对比方法,并且对比相同实验环境下MDNet算法,在线跟踪速度提高了近12倍。

基于自注意力机制的多域卷积神经网络的视觉追踪

为了解决多域卷积神经网络(MDNet)在目标快速移动和外观剧烈变化时发生的模型漂移问题,提出了自注意力多域卷积神经网络(SAMDNet),通过引入自注意力机制从通道和空间两个维度来提升追踪网络的性能。首先,利用空间注意力模块将所有位置上的特征的加权总和选择性地聚合到特征图中的所有位置上,使得相似的特征彼此相关;然后,利用通道注意力模块整合所有特征图来选择性地强调互相关联的通道的重要性;最后,融合得到最终的特征图。此外,针对MDNet算法因训练数据中存在较多相似但属性不同的序列所造成的网络模型分类不准的问题,构造了复合损失函数。该复合损失函数由分类损失函数和实例判别损失函数组成,首先,用分类损失函数来统计分类的损失值;然后,利用实例判别损失函数来提高目标在当前视频序列中的权重,抑制其在其他序列中的权重;最后,融合两项损失作为模型的最终损失。在目前广泛采用的测试基准数据集OTB50和OTB2015上进行实验,结果表明所提出的算法在成功率指标上相比2015年视觉目标跟踪挑战(VOT2015)的冠军算法MDNet分别提高了1.6个百分点和1.4个百分点,在精确率和成功率指标上优于连续域卷积相关滤波(CCOT)算法,在OTB50上的精确率指标优于高效卷积操作(ECO)算法,验证了该算法的有效性。