基于生成对抗机制的多目标跟踪方法

Multi-target Tracking Method Based on Generative Adversarial Mechanism

针对多目标跟踪中的跟踪结果易受遮挡而产生的漏检问题,提出一种公路场景下的基于生成对抗机制的多目标跟踪方法。首先,对经过预训练的跟踪网络输出的特征进行处理,在特征空间中添加自适应的二维遮挡掩码,生成现实生活中难以获得的遮挡样本;其次,利用生成对抗网络在无监督学习方面的优势,将FairMOT模型作为判别网络,加入结合强化学习机制的生成网络来学习如何筛选困难样本,2个网络进行对抗训练以提升多目标跟踪模型的遮挡不变性,提高跟踪精度;最后,在重识别分支中引入中心损失函数以提高重识别准确度。取数据集BDD100K中部分视频序列进行实验,实验结果表明:改进后算法的跟踪准确率提升了0.8个百分点,跟踪精度降低了0.4个百分点,跟踪过程中身份的切换次数减少了208。

A multi-objective tracking method based on generative adversarial mechanism is proposed in highway scenes to address the problem of missed detections caused by occlusion in tracking results in multi-target tracking.Firstly, the features output by the pre-trained tracking network are processed, and adaptive two-dimensional occlusion masks are added to the feature space to generate occlusion samples that are difficult to obtain in real life.Secondly, to leverage the advantages of generative adversarial networks in unsupervised learning, the FairMOT model is used as the discriminative network, and a generative network combined with reinforcement learning mechanism is added to learn how to filter difficult samples.The two networks were trained adversarially to improve the occlusion invariance of multi-target tracking models and the tracking accuracy was improved.Finally, the center loss function was introduced into the re-identification branch to improve the accuracy of re-identification.Experiments were conducted on partial video sequences from the BDD100K dataset.The experimental results show that the improved algorithm improves tracking accuracy by 0.8 percentage points, reduces tracking accuracy by 0.4 percentage points, and reduces the number of identity switches during the tracking process by 208 times.