协同运动状态估计的多目标跟踪算法

Multiple Object Tracking Algorithm via Collaborative Motion Status Estimation

多目标跟踪在视频分析场景中有着广泛的应用,如人机交互、虚拟现实、自动驾驶、视频监控和机器人导航等。多目标跟踪问题可以表示为在已有的检测数据上进行目标轨迹关联,检测算法的准确性对跟踪性能起着关键性的作用。在基于检测的目标跟踪框架中,提出了一种协同运动状态估计的跟踪算法,该算法主要关注相邻帧之间的数据关联,从目标检测、目标运动状态估计和数据关联这3个方面来直接解决多目标跟踪面临的挑战。首先,对于目标检测,采用Multi Scale Convolutional Neural Network(MS-CNN)算法作为检测器,这是因为深度学习在检测的效益上优于传统的机器学习方法;其次,为了更好地预测目标的运动状态和处理目标间的遮挡,针对不同状态的目标采取不同的运动估计方法:采用核相关滤波来评估处于跟踪状态的目标的运动状态,当目标处于遮挡状态时,采用卡尔曼滤波做运动估计;最后,采用Kuhn-Munkres算法对检测目标和跟踪轨迹做数据关联。通过大量的实验证实了算法的有效性,且实验结果表明算法的准确性很高。

Multiple object tracking （MOT） is widely applied in video analysis scenarios, such as human interaction, vir- tual reality, autonomous driving, visual surveillance and robot navigation etc. MOT can be formulated as a sort of track- lets association in existing detection results, in which the accuracy of detection algorithm is entitled an essential role in tracking performance. We proposed a multiple object tracking algorithm via coUaborative motion status estimation. The algorithm is based on the tracking-by-detection framework. The algorithm predominantly focuses on data association of adjacent video frames？tackling challenges of MOT from three aspectsobject detection,object motion status estimation and data association. Firstly,as for object detection, multi scale convolutional neural network（MS^CNN） is adopted as the detector, since the advantage of deep learning in detection outweighs that of classical machine learning method. Se-condly, to better predict object motion status and handle occlusion among targets？ different motion estimation methods are utilized according to different motion statuses. In tracking status, kernelized correlation filter is employed, while in occlusion status,the use of kalman filter is prioritized. Lastly,Kuhn-Munkres algorithm is adopted to work out data as-sociation between detections and tracklets. A substantial amount of experiments were carried out to estimate the effi-ciency. The results are quite positive,demonstrating high accuracy.