一种改进的离散连续能量最小化多目标跟踪

发布时间：2018-01-24 21:33

本文关键词： 机器视觉多目标跟踪能量最小化身份标签互换轨迹分段　出处：《激光与光电子学进展》2017年11期 　论文类型：期刊论文

【摘要】：针对离散-连续能量最小化(DCEM)方法在复杂场景中对轨迹分段或身份标签互换无法有效处理的问题,提出一种改进的DCEM多目标跟踪方法。该方法通过提取被跟踪目标的多特征融合外观向量,利用不同目标间外观特征向量的欧氏距离设计轨迹的外观约束项,处理身份标签互换问题;通过计算相邻时空域内不同轨迹间的运动相似性和外观相似性,设计后处理过程,合并可能为同一轨迹的短轨迹,处理轨迹分段问题。实验结果表明,平均跟踪准确度提高3.6%,平均跟踪精度提高2.5%,并且身份标签互换和轨迹分段情况得到大幅改善,该方法具有更精确更稳定的跟踪能力。
[Abstract]:The discrete continuous energy minimization (DCEM) method can not effectively deal with trajectory segmentation or identity label exchange in complex scenarios. An improved DCEM multi-target tracking method is proposed, in which the appearance vectors of multi-feature fusion are extracted. The Euclidean distance between different objects is used to deal with the problem of identity label exchange. By calculating the motion similarity and appearance similarity between different trajectories in adjacent space-time domain, the post-processing process is designed to merge the short trajectory that may be the same track, and to deal with the trajectory segmentation problem. The average tracking accuracy is increased by 3.6, the average tracking accuracy is improved by 2.5, and the identity label exchange and trajectory segmentation are greatly improved. This method has a more accurate and stable tracking capability.
【作者单位】：江南大学物联网技术应用教育部工程研究中心;
【基金】：中央高校基本科研业务费专项资金(JUSRP51510)
【分类号】：TP391.41
【正文快照】： 多目标跟踪是跟踪图像序列中的多个移动目标并随时间分配各个目标身份的问题,目标图像不完整、遮挡和复杂的目标运动等都会增加跟踪的难度。在发展初期,多目标跟踪主要是独立检测的跟踪,即在第一帧标定被跟踪目标的位置然后用滤波方法逐帧跟踪目标。但这种方法只能跟踪固定数

【相似文献】