图像目标检测前跟踪的广义多伯努利滤波算法

Generalized Multi-Bernoulli filter for track-before-detect of objects from image observations

针对目标影响区域重叠时的图像目标检测前跟踪问题,推导了基于多伯努利滤波器的多目标联合检测与跟踪算法.在分析多个目标叠加条件下观测似然函数的基础上,利用预测得到的目标状态对观测似然函数进行估计,从而消除目标叠加对观测更新带来的影响.该方法在目标预测与跟踪阶段皆保持了目标状态的多伯努利分布特性,是较为严格意义上的多伯努利多目标滤波器,可应用于一般图像观测条件下(目标重叠或非重叠)的目标检测前跟踪.给出了该算法的实现步骤,并通过加标签的方法,更准确地实现目标轨迹提取和虚假目标剔除,最后通过计算机仿真实验验证了所提算法的有效性.

A Generalized Multi-Bernoulli Filter for Track-before-detect（ GMB-TBD） of objects from image observations when the objects＇ influence region overlapping is proposed. The overlapping objects＇ measurement likelihood function is analyzed,the likelihood function is estimated by predicted objects＇ states,and then objects＇ overlapping influence is eliminated on objects＇ states updating by using this estimation. In this filter,the predicted and updated objects＇ states are strictly assumed as Multi-Bernoulli RFS,so it＇s a true Multi-Bernoulli based TBD filter and it can be used under both the objects＇ influence region overlapping and non-overlapping situations. The filter＇s realization steps are given. Objects＇ tracks are pruned and extracted by labeling Multi-Bernoulli components. Lastly,GMB-TBD filter＇s performance is verified by computer Monte-Carlo simulation results.