计算机与现代化

• 算法分析与设计 • 上一篇    下一篇

基于参数学习和运动预测的在线多目标跟踪算法

  

  1. 1.中国科学院电子学研究所,北京100190;2.中国科学院大学,北京100049
  • 收稿日期:2017-02-22 出版日期:2017-10-30 发布日期:2017-10-31
  • 作者简介:李鹏飞(1992-),男,河北邢台人,中国科学院电子学研究所、中国科学院大学硕士研究生,研究方向:计算机视觉,多目标跟踪; 雷宏(1963-),男,中国科学院电子学研究所研究员,博士生导师,研究方向:电磁场与微波技术,信号处理理论与技术。

Online Multiple Object Tracking Based on Parameter Learning and Motion Prediction

  1.  
    1. Institute of Electronics, Chinese Academy of Sciences,Beijing 100190, China;
     2. University of Chinese Academy of Sciences, Beijing 100049, China
  • Received:2017-02-22 Online:2017-10-30 Published:2017-10-31

摘要: 针对在线多目标跟踪中的短时遮挡和检测器误差造成的误检和漏检问题,提出一种结合参数学习和运动预测的在线多目标跟踪算法。采用逐帧关联的方式,首先利用目标的历史轨迹建立卡尔曼滤波器模型预测目标当前帧的可能位置,然后计算目标和当前观测之间的关联度建立代价矩阵。对于多目标跟踪被建模为指派问题,采用Hungarain算法求解,此外制定策略处理目标出现、消失和遮挡等异常情况。而对于多目标跟踪系统中的参数,设计一种新的二分类参数学习方案。实验结果验证了参数学习的有效性以及对误检、漏检和遮挡的鲁棒性,并且与若干经典算法的性能比较中,在多个指标上表现出一定优势。

关键词: 在线多目标跟踪, 卡尔曼滤波, 指派问题, 参数学习, 运动预测

Abstract: For short term occlusion and detector errors in online multiple object tracking, a new algorithm based on parameter learning and motion prediction is proposed. Firstly, the Kalman filter model is established by using the historical trajectory of the target, and target possible position in the current frame is given. Then, the cost matrix is established by calculating the correlation between the target and the current observation. The multi-target tracking is modeled as an assignment problem, and the Hungarain algorithm is used to solve the problem. In addition, the unusual situation of the target entering, disappearing and occlusion are processed. For the parameters of the multi-target tracking system, a new binary classification learning scheme is designed. Experimental results verify the effectiveness of parameter learning and the robustness against false detection, missed detection and occlusion. The proposed method has some advantages in many aspects compared with the performance of several classical algorithms.

Key words: online multiple object tracking, Kalman filter, assignment problem, parameter learning, motion prediction