基于联合注意力孪生网络目标跟踪算法

Object tracking algorithm based on siamese network with combined attention

为改进在发生形变、尺度变化及相似目标等多种干扰因素时视频中运动目标的跟踪精度,提出了一种联合注意力的孪生网络模型。首先,采用一种轻量级网络MobileNetV3作为主干网络对目标进行特征提取;然后,为提高模型对于目标关键特征的关注度,提出了通道联合空间注意力与孪生网络结合的模型结构;最后,对基于注意力模块与非注意力模块的特征向量互相关结果进行加权融合获得响应图,并利用该响应图获得目标跟踪结果。实验结果表明,所提算法在OTB50与OTB100数据集上能够获得较好的跟踪效果,两个数据集平均精确率和成功率达到78.5%和58.3%。此外,当存在形变、尺度变化及相似目标等不合作因素时,所提算法仍能取得较好的跟踪效果,从而表明该算法具有良好的鲁棒性。

In order to improve the tracking accuracy of moving targets in video when various interference factors such as deformation, scale variation and similar targets occur, a siamese network model with combined attention is proposed. Firstly, a lightweight network, i.e., MobileNetV3, is adopted as the backbone network to extract object feature. Then, in order to improve the attention of the model to the key features of the target, a model structure combining channel combined spatial attention and siamese network is proposed. Finally, through weighting and fusing the cross-correlation results of the feature vectors of attention module and non-attention module, the response map can be obtained, which can be used to obtain the tracking result. Experiment results show that the proposed algorithm can achieve good tracking effect on the OTB50 and OTB100 datasets, the average accuracy and success rate for the two datasets reach 78.5% and 58.3%, respectively. In addition, when multiple uncooperative factors, such as deformation, scale variation and similar targets exist, the proposed algorithm can still achieve good tracking effect, which shows that the proposed algorithm has good robustness.