时序姿态扰动补偿下无人机多模态特征混叠目标视觉标定系统设计

摘要:

无人机在持续飞行任务中因温度变化与机械振动会使得视觉传感器参数漂移,导致多模态数据时序失准与特征混叠等动态感知失调问题。鉴于以上问题,设计时序姿态扰动补偿下无人机多模态特征混叠目标视觉标定系统。系统通过构建多模态传感协同采集、FPGA高精度同步触发与姿态-传感联动补偿的硬件架构,实现视觉、惯性、位置数据的时空对齐与动态姿态抑制。在特征层面,提出改进的分区加权LBP纹理建模方法,结合时序-姿态联合补偿机制,实现多源底层特征的深度融合与偏差修正,有效解耦因振动与异步导致的特征混叠。变电站绝缘子缺陷检测实验表明:系统在污秽识别中LBP直方图方差全部落入≤70的理想区间,伞裙裂纹三维重建长度误差仅为0.5cm,显著优于传统方法,验证了系统在复杂动态场景下的高精度与强鲁棒性。

Abstract:

During continuous flight missions of unmanned aerial vehicles (UAVs), temperature changes and mechanical vibrations can cause the parameters of visual sensors to drift, leading to dynamic perception disorders such as inaccurate timing of multimodal data and feature aliasing. In view of the above problems, a multi-modal feature aliasing target visual calibration system for unmanned aerial vehicles under temporal attitude disturbance compensation is designed. The system achieves spatio-temporal alignment and dynamic attitude suppression of visual, inertial and position data by constructing a hardware architecture featuring multi-modal sensing collaborative acquisition, high-precision synchronous triggering by FPGA and attitude-sensing linkage compensation. At the feature level, an improved partitioned weighted LBP texture modeling method is proposed, combined with the temporal-pose joint compensation mechanism, to achieve deep fusion and deviation correction of multi-source low-level features, effectively decoupling feature aliasing caused by vibration and asynchrony. The experiments on insulator defect detection in substations show that the variance of the LBP histogram of the system in pollution identification all fall within the ideal range of ≤70, and the length error of the three-dimensional reconstruction of the umbrella skirt crack is only 0.5cm, which is significantly better than the traditional method, verifying the high precision and strong robustness of the system in complex dynamic scenarios.