深度学习辅助的多行人跟踪算法

发布时间：2018-05-13 13:06

  本文选题：多目标跟踪 + 识别辅助的跟踪　； 参考：《中国图象图形学报》2017年03期

【摘要】：目的目标的长距离跟踪一直是视频监控中最具挑战性的任务之一。现有的目标跟踪方法在存在遮挡、目标消失再出现等情况下往往会丢失目标,无法进行持续有效的跟踪。一方面目标消失后再次出现时,将其作为新的目标进行跟踪的做法显然不符合实际需求;另一方面,在跟踪过程中当相似的目标出现时,也很容易误导跟踪器把该相似对象当成跟踪目标,从而导致跟踪失败。为此,提出一种基于目标识别辅助的跟踪算法来解决这个问题。方法将跟踪问题转化为寻找帧间检测到的目标之间对应关系问题,从而在目标消失再现后,采用深度学习网络实现有效的轨迹恢复,改善长距离跟踪效果,并在一定程度上避免相似目标的干扰。结果通过在标准数据集上与同类算法进行对比实验,本文算法在目标受到遮挡、交叉运动、消失再现的情况下能够有效地恢复其跟踪轨迹,改善跟踪效果,从而可以对多个目标进行持续有效的跟踪。结论本文创新性地提出了一种结合基于深度学习的目标识别辅助的跟踪算法,实验结果证明了该方法对遮挡重现后的目标能够有效的恢复跟踪轨迹,适用在监控视频中对多个目标进行持续跟踪。
[Abstract]:Target long-range tracking is one of the most challenging tasks in video surveillance. The existing methods of target tracking often lose the target in the presence of occlusion and disappear and reappear, so they can not be tracked continuously and effectively. On the one hand, tracking a target as a new target when it reappears after disappearing is clearly not in line with actual needs; on the other hand, when a similar target appears in the tracking process, It is also easy to mislead the tracker to treat the similar object as a tracking target, resulting in tracking failure. Therefore, a tracking algorithm based on target recognition assistance is proposed to solve this problem. Methods the tracking problem is transformed into the problem of finding the corresponding relationship between the detected targets between frames. After the target vanishes and reappears, the depth learning network is used to achieve effective trajectory recovery and to improve the effect of long distance tracking. And to a certain extent to avoid the interference of similar targets. Results by comparing with the similar algorithms on the standard data set, the algorithm can recover the tracking track effectively and improve the tracking effect when the target is occluded, cross moving, disappear and reappear. Thus, multiple targets can be tracked continuously and effectively. Conclusion in this paper, a target recognition aided tracking algorithm based on depth learning is proposed. The experimental results show that this method can effectively recover the track of the target after occlusion reconstruction. Suitable for continuous tracking of multiple targets in surveillance video.
【作者单位】： 浙江工商大学计算机与信息工程学院;北京正安维视科技股份有限公司;兰州大学信息科学与工程学院;
【基金】：国家自然科学基金项目(61472362,61379075) 浙江省自然科学基金项目(LZ16F020002,LY14F020001) 公益技术研究社会发展项目(2015C33081)~~
【分类号】：TP391.41
，

本文编号：1883270