一种融合运动特征嵌入的多目标分割跟踪算法

Multi-object Tracking and Segmentation Algorithm by Fusing Motion Feature Embedding

针对现有多目标跟踪算法中存在目标运动模糊和相互遮挡的难点,在单阶段和无锚框的实例分割框架下,提出了一种融合运动特征嵌入的多目标分割跟踪算法.首先,提取当前帧与前后两帧光流场中的运动信息对表观特征进行运动补偿,再利用特征金字塔网络融合含有运动信息的多尺度特征,提高了目标检测性能.其次,通过两个用于提升网络预测性能的损失函数的设计和使用,进一步减少了由于检测器失效和目标遮挡而导致的漏检.最后,关联网络提取目标的外观特征,并通过预测并关联的更新轨迹策略将可靠的跟踪结果合并至轨迹.实验结果表明,本文提出的算法在MOTS20训练集上跟踪准确度达到了66.0%,测试集上达到了63.1%,与同类算法相比,本文算法表现出更好的有效性.

In order to solve the problems of motion blur and mutual occlusion in existing multi-target tracking algorithms,this paper proposes a multi-object tracking and segmentation algorithm by fusing motion feature embedding under a single-stage and anchor-free detection framework.Firstly,appearance features are compensated using motional cues through estimating the optical flow field between current and neighboring frames,and feature pyramid network is used to fuse multi-scale features containing motion information,which effectively improves performance of detecting objects.Secondly,two loss functions are designed to improve network prediction performance,which further reduce false negative due to detector failure and object occlusion.Finally,association network extracts the appearance characteristics of the targets and the predicted and associated update trajectory strategy merges reliable tracking results into the trajectory.Experimental results show that the tracking accuracy of proposed method reaches 66.0%on MOTS20 training set and 63.1%on test set.Compared with methods using similar frameworks,the proposed method shows better effectiveness.