基于多信息融合的多目标跟踪方法研究

Multi-target Tracking Method Based on Multi Information Fusion

在线多目标跟踪作为计算机视觉和人工智能方面的一个研究热点,随着深度学习的发展取得了较大的进展;但是依然存在诸如复杂场景跟踪准确度低等亟待解决的问题;针对多目标跟踪研究中存在的行人特征信息较少、跟踪目标被遮挡等问题,提出了一种融合表观信息、轨迹历史信息和目标运动信息的多目标跟踪方法,通过专门设计的双分支网络结构和损失函数使模型在学习时将三种信息相互融合;改进相似性分数计算方法获得更多的特征信息,提取更为鲁棒的特征;多信息融合的多目标跟踪方法在计算方面开销较少,能够在测试时达到实时的效果;并且,通过相关实验验证,基于多信息融合的多目标跟踪方法能够在MOT16数据集上达到很好的性能,可以更好地处理目标遮挡、目标误检及目标丢失等情况。

As a research hotspot in computer vision and artificial intelligence,online multi-target tracking has made great progress with the development of deep learning in recent years.However,there are still many problems to be solved,such as low tracking accuracy of complex scenes.Aiming at the problems of pedestrian feature information in multi-target tracking research and occlusion of tracking targets,a multit-arget tracking method is proposed to combine apparent information,trajectory historical information trajectory historical information and target motion information.The designed two branch network structure and loss function enable the model to fuse these information with each other during learning to obtain more feature information and extract more robust features,which improves the similarity score calculation method to obtain more feature information and extract more robust features.The proposed method has less computational overhead and can achieve realtime effects during testing.In addition,relevant experimental results show that the proposed method can achieve good performance on MOT16 and other data sets,which can better deal with target occlusion,target misdetection,and target loss.