基于改进的Transformer加Anchor-free网络的多目标跟踪算法

Improved Transformer Plus Anchor-Free Network Based on Multi-Object Tracking Algorithm

近年来,基于Anchor-free的多目标跟踪算法以其精度高、速度快、超参数少的特点被广泛研究。但是,实际场景中的目标遮挡使得此类算法仍然面临挑战,这类算法会对遮挡后重新出现的目标的身份信息进行错误切换。针对以上问题,提出了一种基于改进的Transformer加Anchor-free网络的多目标跟踪算法(Transformer-Anchor-free-MOT,TransAnfMOT),该算法通过跨层特征融合(Cross-Layer Feature Fusion, CFF)和注意力机制(Convolutional Block Attention Module, CBAM)对RGB图像和Depth图像进行融合,提升融合的RGB-D图像的特征质量,从而提高遮挡判断任务的精度。另外,设置了被遮挡目标的搜索区域,并利用外观特征距离方法赋予遮挡后重新出现的目标之前的身份信息,减少目标身份信息切换错误。实验结果表明,提出的算法在3个不同的场景下实现了比较有竞争力的结果,有效提升了多目标跟踪算法的准确性和稳定性。

In recent years,Anchor-free based Multi-Object Tracking(MOT)algorithms have been widely studied for their high accuracy,speed and few hyperparameters.However,the object occlusion in real-world scenarios still make this kind of algorithm challenging.Such Anchor-free based on MOT algorithms can incorrectly switch the identity information of objects that reappear after occlusion.To address the above problems,an improved Transformer plus Anchor-free network is proposed based on Multi-Object Tracking Algorithm Transformer-Anchor-free-MOT(TransAnfMOT),which fuses RGB and Depth images through cross-layer feature fusion(CFF)and the convolutional block attention module(CBAM)to enhance the feature quality of fused RGB-D images and improve the accuracy of occlusion judgement tasks.In addition,the search area for the occluded objects is set and the appearance feature distance method is used to assign the previous identity information to the objects that reappear after occlusion,which reduce the object identity information switching errors.Experimental results show that the proposed algorithm achieves more competitive results in three different scenarios,effectively improving the accuracy and stability of the multi-object tracking algorithm.