基于弱监督贝叶斯推断模型的多无人机协同跟踪技术研究

Due to the uncertainties raised from both the application environment and the motion of the objects, it is difficult to maintain robust and continues tracking by using a single unmanned aerial vehicle (UAV), especially in the presence of object occlusions. Deployment of multiple UAVs allows for observation of the objects from different point of views, which is a nature solution to the occlusion problem in a visual tracking system. In order to achieve cooperation across UAVs, it is usually required to establish the spatial-temporal constraints of the targets moving across the field of view for each UAV, and then solve the associated correspondence problem. In this work, we present a general framework to simultaneously estimate the motion distribution of the objects and infer the correspondences of the target through the Bayesian Inference method. In addition, we further develop a single bag multiple instance learning (MIL) method, which allows predicting instance labels by using only positive bags. This newly developed MIL is applied to solve the Bayesian inference model training problem, which improves the prediction accuracy when compared with commonly used unsupervised training scheme. The efficiency and flexibility of the proposed techniques would be demonstrated in simulated scenarios using multiple small UAV platforms.

由于目标运动的不确定性和应用环境的复杂性，使用单无人机执行目标跟踪任务时无法完全突破由于目标遮挡所导致目标丢失的技术瓶颈。通过多无人机的协同能够获取多个视角的目标状态信息，是解决目标遮挡问题的有效手段。为实现对目标的协同跟踪，首先要建立目标在多个无人机视域内的运动分布模型以及目标转移的时空约束，然后以此为条件求解目标在多无人机视域中的匹配问题。针对这两类子问题彼此相互关联但单独求解困难的问题特性，本课题基于“共治”思想，将提出一种基于贝叶斯推断理论的求解策略，达到同时估计目标运动分布模型和关联目标标记的目的。为解决由于目标和多无人机之间机动所引起的模型参数变化问题，本课题拟提出一种仅依靠“正包”样本的多示例学习机制，能够利用自动生成的训练集实现推断模型参数的在线估计。依托课题组所在的教育部工程研究中心的软硬件资源，拟开展模拟环境中进行多无人机协同跟踪的试验从而验证预期研究成果的有效性。

随着无人机技术的飞跃式发展, 利用无人机跟踪地面目标在军事和民用领域得到了广泛的应用. 鉴于目标所在环境以及运动状态的不确定性, 单一无人机已不能胜任日益复杂的应用环境, 利用多无人机协同跟踪目标成为改善目标跟踪任务鲁棒性的一种有效手段. 本项目针对多无人机协同跟踪问题开展研究工作, 分析了目标在多个无人机视域内的运动分布模型以及目标转移的时空约束，基于“共治”思想提出一种基于贝叶斯推断理论的求解策略，结合目标与无人机间的时空约束对目标进行有效定位和跟踪。针对目标在无人机视域中特征变化大、特征选择困难的问题，本项目基于多示例学习机制，提出了一种仅依靠“正包”样本的半监督学习算法，能够利用自动生成的训练集实现推断模型参数的在线估计。本项目所研究的协同跟踪方法在无人机协同目标指示与任务分配演示验证项目中得到了应用，结果表明本项目所形成的成果能够很好的解决多机协同目标跟踪问题，具有很好的应用前景。