【摘要】 在视频的多目标跟踪任务中，卡尔曼滤波器性能受硬件噪声以及光线等环境噪声干扰较大，导致滤波性能下降甚至发散，严重影响目标跟踪精度。针对这一问题，在检测端不变的情况下，对跟踪算法中的卡尔曼滤波器进行改进。首先，通过实时监测跟踪过程中滤波器观测值和估计值的动态变化，提取新息或残差；然后，利用新息协方差对观测噪声统计特性进行自适应估计，进而调整卡尔曼滤波增益；并通过数值仿真表明所提方法能有效降低噪声，获得更好跟踪效果。最后，基于YOLOv3算法检测结果进行实验验证，结果表明在多目标跟踪（MOT16）数据集上，相较于传统卡尔曼滤波设计，所提自适应卡尔曼滤波在多目标跟踪任务中的精度、标号（ID）相关指标（IDF1,IDP）等指标均有所提升。更多还原

【Abstract】 In the tasks of multiple object tracking,performance of filtering is deteriorated due to the fact that Kalman filter is sensitive to the noise from hardware device and environment,which leads to a decrement on the tracking precision.To address the adverse effect of noise,an improvement of the Kalman filter in the tracking algorithm was proposed with the detector unchanged. Firstly,innovations or residuals were extracted by real-time surveillance on the dynamic variation of filter observations and estimations. Then,the statistical characteristic of the observed noise was estimated adaptively by the covariance of innovations to adjust the gain of Kalman filter;and a simulation experiment demonstrates the validity of the proposed method in terms of restraining noise to obtain better tracking performance. Experimental results on multiple object tracking dataset MOT16 by using YOLOv3 for detection show that the proposed adaptive Kalman filter outperforms other multiple object tracking algorithms based on classical Kalman filter in precision and ID indicators(IDF1,IDP)as well as other evaluation indicators.更多还原