复杂环境下基于多视角与轻量级卷积神经网络的多无人机目标跟踪

Vision-based object tracking of the unmanned aerial vehicle (UAV) in the complex environment is a very challenging task. This project aims to develop an object tracking method for multiple UAVs based on multiple views and lightweight convolutional neural network. Aiming at the convolution neural network with complex structure adopted by the existing object tracking methods, an onboard lightweight convolution neural network which is applicable for UAV platform will be researched. Considering the advantage of multiple views of multiple UAVs, the approach to train and optimize the object model with enough valid shared samples will be explored, improving the robustness and accuracy of object tracking in the complex environment. Aiming at the limitation of existing hand-crafted feature-based UAV object tracking method, an object tracking method by online selecting convolutional layer dynamically will be researched, developing a UAV object tracking approach based on convolutional features. Aiming at many uncertain factors in the complex environment, a fuzzy control approach with non-singleton dynamic fuzzification and Gaussian model will be explored, controlling the UAV with high accuracy and speed. The research results of this project are significant to solve the existing critical problems in the current research of object tracking of the UAV. This project is also beneficial for enriching the research theories of UAV tracking and promoting the object tracking of the UAV.

基于视觉的无人机目标跟踪在复杂环境下是一项非常具有挑战性的任务。本项目拟针对基于多视角与轻量级卷积神经网络的多无人机目标跟踪方法开展研究。针对现有目标跟踪方法采用结构复杂的卷积神经网络特点，研究轻量级卷积神经网络的构建方法，以适用无人机的机载平台使用；利用多无人机具有多视角的优势，研究通过共享足够有效样本对目标模型进行归一化训练与优化的方法，以提高复杂环境下目标跟踪的鲁棒性与准确性；针对现有无人机机载目标跟踪方法采用手工设计特征的局限性，研究基于在线动态选取卷积层的目标跟踪方法，进而提出基于卷积特征的无人机机载目标跟踪方法；针对多样化复杂环境中往往存在诸多不确定性因素的问题，研究非单点动态高斯型模糊控制方法，以实现对无人机的快速精确控制。预期成果将为复杂环境下基于视觉的无人机目标跟踪研究提供理论指导，突破无人机在跟踪性能提升方面所存在的关键问题，对推动无人机目标跟踪的发展具有重要意义。

本项目以多无人机系统为研究对象，通过大规模定性定量分析与实验验证，提出了基于多视角与轻量级卷积神经网络的多无人机目标跟踪方法的关键基础理论。具体地，针对现有目标跟踪方法采用结构复杂的卷积神经网络的缺点，构建了适用于无人机机载平台使用的轻量级卷积神经网络；为提高复杂环境下目标跟踪的鲁棒性与准确性，利用多无人机具有多视角的特点，通过共享足够有效的样本，形成了对目标模型进行归一化训练与优化的方法；针对现有无人机机载目标跟踪方法采用手工设计特征的局限性与采用卷积特征的优势，提出了基于在线动态选取卷积层的目标跟踪方法；为保证多无人机在执行自主跟踪任务中的综合控制性能，针对多样化复杂环境中往往存在诸多不确定性因素的问题，形成了非单点动态高斯型模糊控制方法。上述成果不仅为基于视觉感知的无人机自主目标跟踪提供理论和方法指导，亦对推动无人机目标跟踪进一步发展、促进其在军事与民用领域应用具有极为重要的科学价值与工程实践意义。