概率扩充和改进OIM损失的多目标跟踪算法
Multi-object tracking algorithm of probability expansion and improved OIM loss
付小珊 1胡乃平 1秦建伟 1王传旭1
作者信息
- 1. 青岛科技大学信息科学技术学院,山东青岛 266061
- 折叠
摘要
为解决多目标跟踪中联合目标检测和重识别训练时间过长、多分支特征不对齐和目标相互遮挡的身份转换问题,提出一种高效的多目标跟踪算法.在特征提取阶段利用深层聚合网络联合多层次特征,在重识别阶段通过三元组对在线实例匹配损失进行增强,缓解特征不对齐问题.加入高斯核函数对训练样本进行概率扩充,缩短训练时间.利用运动、外观特征与卡尔曼滤波实现高效的在线关联,利用轨迹池暂存丢失的轨迹,提高目标相互遮挡时的跟踪性能.算法在MOT15和MOT17数据集上的准确度分别达到了 60.1％与74.2％,MOT17上的FPS也达到21.6 Hz.
Abstract
To solve the problems of long training time for joint object detection and re-identification,multi-branch feature misa-lignment,and identity transformation with mutual object occlusion in multi-object tracking,an efficient multi-object tracking algorithm was proposed.The feature misalignment problem was mitigated using a deep aggregation network to combine multi-level features in the feature extraction stage and the online instance matching loss was augmented by triples in the re-recognition stage.A Gaussian kernel function was added to probabilistically expand the training samples to shorten the training time.Motion and appearance features were used with Kalman filtering to achieve efficient online association,and a trajectory pool was used to temporarily store the lost trajectories to improve the tracking performance when targets were occluded from each other.The algorithm achieves 60.1％and 74.2％accuracy on MOT15 and MOT17 datasets,respectively,and 21.6 Hz FPS on MOT17.
关键词
多目标跟踪/目标检测/重识别/深层聚合/高斯核/在线实例匹配/卡尔曼滤波Key words
multi-object tracking/object detection/re-identification/deep layer aggregation/Gaussian kernel/online instance matching/Kalman filtering引用本文复制引用
出版年
2024
NETL
NSTL
万方数据