融合HOG特征和注意力模型的孪生目标跟踪算法

为了提高跟踪算法在目标发生形变和被遮挡时的准确性,提出一种融合HOG(histogram of oriented gradient)特征和注意力模型的孪生目标跟踪算法.首先,采用对ResNet残差模型改进后的CIR(cropping inside residual)模型塑造孪生目标跟踪网络的骨干网络,充分利用不同层次的特征图,同时加深网络;其次,融入HOG特征,增强网络对图形几何变化的鲁棒性;再次,加入CBAM(convolutional block attention module)注意力模型,使网络能够在结合上下文信息的同时调节HOG特征在特征图中所占比例,增强特征图中的有效特征,弱化无效特征,使网络中各特征图发挥出最好的效果;最后,定义算法的损失函数.实验结果表明,所提算法在GOT-10k数据集上进行训练后,能够在OTB100上获得较好的跟踪效果,在该数据集中精确率和成功率分别达到81.9%和60.6%.在目标物体发生形变和被遮挡的情况下,所提算法仍能取得较好的跟踪效果.

Twin target tracking network combining HOG features and attention model

In order to improve the accuracy of the tracking algorithm when the target is deformed and occluded, a twin target tracking algorithm integrating a HOG(histogram of oriented gradient,HOG) feature and an attention model is proposed . First, the CIR model improved by the ResNet residual model is used to shape the backbone network of a twin target tracking network, which making full use of different levels of feature maps to deepen the network. Secondly, the HOG feature is integrated to enhance the robustness of the network to the geometric changes of graphics. Thirdly, the CBAM(convolutional block attention module) attention model is added to enable the network to adjust the proportion of HOG features in the feature map while combining the context information, enhance the effective features in the feature map, weaken the invalid features, and make each feature map in the network play the best effect. Finally, the loss function of the algorithm is defined. Experimental results show that the proposed algorithm can achieve good tracking effect on OTB100 after training on GOT-10k dataset, and the accuracy and success rates in this dataset are 81.9% and 60.6%, respectively. When the target object is deformed and occluded, the proposed algorithm can still achieve better tracking results.