厚尾噪声环境下基于置信传播的鲁棒跟踪算法

Robust Tracking Algorithm Based on Belief Propagation in Heavy-Tailed Noise Environments

多目标跟踪技术可从目标个数未知且存在漏检、杂波和噪声的复杂跟踪环境中同时估计出目标状态和目标个数,已经广泛应用于空中预警、自动驾驶和移动机器人等领域,对国防军事和民生科技都具有重要的应用价值。然而,在实际的跟踪环境中,外界的干扰和传感器自身的不稳定会使得量测噪声中出现野值而表现出厚尾特性,此外,目标在杂波环境下发生机动时,运动模型的不准确也会产生厚尾过程噪声。此时若继续在高斯假设下进行多目标滤波处理会使得跟踪精度大大下降。针对该问题,常用的解决方案是将厚尾的过程噪声和量测噪声建模为学生t分布,并用来修正随机有限集(Random Finite Set,RFS)理论下的标准多目标滤波器,进而确保跟踪性能不发散。然而,基于RFS理论的多目标跟踪方法往往需要较大的计算代价,导致系统的延迟增加。本文利用扩展性强且计算复杂度较低的置信传播(Belief Propagation,BP)策略,提出了一种基于BP的多目标鲁棒跟踪算法。该算法首先将每个目标的后验概率密度函数近似为学生t分布混合模型,然后通过BP模式进行递归更新,最后基于判决门限,实现目标状态的估计。仿真实验表明,相比于现有算法,本文提出的算法能够在过程噪声和量测噪声同时存在厚尾时的跟踪场景中实现稳健有效的跟踪性能。

Multi-target tracking technology can simultaneously estimate the states and quantities of targets in complex tracking environments where the number of targets is unknown and missed detections,clutter,and noise exist.This technology has been widely applied in fields such as airborne early warning,autonomous driving,and mobile robotics,with significant application value in both defense and civilian technologies.However,in practical tracking environments,external interference and sensor instability can lead to outliers in measurement noise,exhibiting heavy-tailed characteristics.Additionally,the inaccurate motion models of targets in cluttered environments can generate heavytailed process noise.Continuing multi-target filtering under the Gaussian assumption in such scenarios significantly reduces tracking accuracy.A common solution to addressing this issue is to model heavy-tailed process and measurement noise as Student’s t-distributions and use them to correct standard multi-target filters under the random finite set(RFS)theory,thereby ensuring that the tracking performance does not diverge.However,multi-target tracking methods based on the RFS theory often incur substantial computational costs,resulting in increased system latency.This study proposes a robust multi-target tracking algorithm based on belief propagation(BP),leveraging its strong scalability and low computational complexity.The algorithm first approximates the posterior probability density functions of each target as Student’s t-distribution mixture models,then recursively updates them through BP iterations,and finally estimates target states based on decision thresholds.Simulation experiments demonstrate that,compared to existing algorithms,the proposed algorithm achieves robust and effective tracking performance in scenarios with heavy-tailed processes and measurement noise.