面向弱特征多目标的车载激光平台快速高精度稳定指向研究

With the popularization of UAVs in various fields, characteristics that UAVs can be equipped with functional equipment casually and are subject to subjective awareness make it to be a serious threat to public safety. Development of automotive laser defense systems against UAVs becomes the main development trend. This topic is confronted with demands of detection, tracking and targeting of low-altitude non-cooperative UAVs. As to the problem of multi-target detection with weak feature under the low-altitude complex environment, multi-target tracking, and fast and high-precision laser beam aiming under carrier interference, it is focused on solving the critical scientific problems of construction of adjacent layer feature fusion based enhanced SSD network model and design of visual reasoning model based on relationships between targets, training set expansion under virtual environment and hyper-parameter optimization based on group intelligence and bayesian inference, and adaptive composite control of photoelectric platform fused with neural network under unknown disturbance. Research is carried out on aspects of detection and tracking of multi targets with low-altitude weak features, data enhancement and acceleration training methods for deep network models, high-precision stabilization of laser beam on the vehicle platform with complex disturbance, as well as high-performance motor drive frame for photoelectric platform and integration development of embedded system. The final goal is to realize design of vehicle-mounted photoelectric platform systems with characteristics of timely accurate detection of low-altitude non-cooperative UAVs and high-precision fast targeting. Meanwhile, it can further improve design theory of intelligent detection and tracking algorithms for photoelectric platform multi-target and high-performance miniaturized photoelectric servo system.

随着无人机在各领域的普及应用，可搭载功能设备且受人主观意识影响的特点使得无人机平台对公共安全造成严重威胁，开发针对无人机平台的车载激光防御系统成为主要发展趋势。本课题面向低空非合作无人机检测、跟踪和瞄准等需求，针对低空复杂环境下弱特征多目标检测，多目标在线跟踪问题，重点解决基于邻近层特征融合SSD深度网络模型和目标间关系可视化推理模型的构建，面向虚拟环境训练集扩充和基于群智能和贝叶斯深度网络超参数优化方法，以及未知干扰下基于融合网络的光电平台复合控制策略等科学问题，对面向低空弱特征多目标检测跟踪算法，深度网络模型数据增强和加速训练手段，大移动干扰条件下激光光束高精度稳向，光电平台高性能驱动框架与嵌入式系统集成开发几个方面开展研究，实现针对低空非合作无人机的及时准确检测和具备高精度快速瞄准功能的车载光电系统研制，进一步完善光电平台多目标智能识别及跟踪算法和高性能小型化稳定伺服系统的设计理论。

车载激光防御系统凭借其反应速度快、精度高、作战效费比高等优点，在低空防御系统中具有重要应用。本项目围绕激光防御系统的弱特征多目标检测跟踪和激光光束高精度稳定指向及快速切换两个方面问题开展研究。具体针对增强SSD深度网络的弱特征多目标检测跟踪算法、目标检测训练集扩充和加速训练研究、车载平台粗精环快速高精度稳定控制研究，平台高带宽驱动设计及嵌入式系统集成开发四个方面开展研究。为提高SSD网络对弱小目标的检测精度，采用稠密特征提取网络替换主干特征提取网络，并引入分组卷积和Focal Loss损失函数。为提高Deepsort跟踪算法的鲁棒性，引入DIoU作为度量标准，采用时间序列加权的方式引入目标历史外观特征。基于Tensorflow框架构建目标检测网络，采用英伟达显卡进行深度网络训练，基于Zynq嵌入式平台对目标检测网络进行硬件加速，从而实现低空弱特征小目标的快速高精度检测。在平台高精度稳定控制方面，设计了基于双速度环的扰动观测器方法，该方法可以提高系统的响应速度，增强系统的扰动抑制能力。另外，针对光电平台在低速转动时受摩擦力影响较大的问题，通过最小二乘法对摩擦模型静态参数进行分段拟合，采用智能差分进化算法辨识摩擦模型动态参数，从而实现对光电平台的平稳低速运行。在光电平台驱动设计和嵌入式系统集成方面，提出了一种基于Zynq嵌入式平台和改进式SVPWM的驱动控制方法，极大地缩减了系统的闭环时间，实现了高带宽、低纹波的驱动模块设计。同时，利用MATLAB/Simulink工具搭建了一套基于新型两相混合式步进电机、SVPWM驱动控制以及三闭环控制的高精度稳定跟踪仿真模型，为针对性的进行驱动控制设计工作提供了有效手段。通过以上研究，形成了激光防御系统中目标检测跟踪、高精度稳定跟瞄方面的理论和实验方法，为实现低空防御装备的高性能研制提供了基础。