目的 基于卡尔曼滤波的视频目标跟踪算法需要事先获得过程噪声和观测噪声方差,但在实际应用中,无法得知这两种噪声方差的准确值。此外,由于目标运动的随机性和视频场景中背景的复杂性,噪声方差也会随时间发生动态变化。如果设定的噪声方差不准确,跟踪精度会受影响,严重时会导致目标跟踪失败。考虑到上述问题,提出一种新的解决方法。方法 将带遗忘因子的推广递推最小二乘法(EFRLS)运用到视频目标跟踪研究领域。在该算法中,无需使用噪声方差,首先利用Mean Shift算法获得目标位置的初步估计,再利用EFRLS算法估计下一帧目标的位置。结果 该算法明显好于传统Mean Shift算法,并且与Kalman结合Mean Shift算法的跟踪性能相当。此外,在目标发生严重遮挡时,该算法优于Kalman结合Mean Shift算法,具有较好的跟踪性能。结论 本文算法无需设置噪声参数,可以实现目标在发生严重遮挡和遮挡后目标重新出现的情况下的准确跟踪,提高了跟踪的鲁棒性,具有一定的工程使用价值。

Objective Video target tracking algorithms based on Kalman filter require prior information, such as process noise and observation noise variance. However, we cannot determine the exact values of both noise variance in practical applications. Moreover, noise variance occasionally changes dynamically because of target randomness and background video scene complexity. If noise variance is inaccurate, tracking accuracy is degraded or tracking failure is brought out. In view of these problems, a new solution is proposed in this paper. Method In combination with the forgetting factor recursive least squares (EFRLS) method, a new algorithm that is applied to video target tracking without use of noise variance is presented in this article. First, mean shift is used to obtain a preliminary estimate of the target position. Then, the EFRLS method is used to estimate the position in the next frame. Result Experimental results show that the proposed algorithm is significantly better than traditional mean shift algorithm and is equivalent to Kalman tracking algorithm combined with mean shift. In addition, if severe occlusions exist in between targets, this algorithm is better than Kalman tracking algorithm combined with mean shift. The proposed algorithm also has good tracking performance. Conclusion Setting parameters of noises is not necessary. Accurate tracking results can be achieved when serious occlusions exist or re-emerging occurs after occlusions. The robustness of the new algorithm is then enhanced, which can be used for some engineering applications.