无人机弱纹理与动态场景下的视觉惯性里程计系统优化方法研究

Research on the Optimization Method of Visual Inertial Odometer System in Weak Texture and Dynamic Scene of Unmanned Aerial Vehicle

摘要: 针对无人机在GPS信号丢失或不可用的复杂环境(如室内、夜间、动态场景)中的自主定位导航需求，提出并实现了一套高性能视觉惯性里程计(VIO)系统的优化方法。系统融合视觉传感器与惯性测量单元(IMU)数据，获取最终无人机位姿信息，然而传统视觉惯性里程计方法易受弱纹理、动态光照、非高斯噪声与外观变化等影响。该方法在图像前端采用CLAHE局部自适应直方图均衡化与Gamma矫正算法进行图像增强处理，有效提升弱光与低纹理场景下的图像质量；同时，引入GFTT + BRISK组合特征提取策略，提高特征稳定性与匹配鲁棒性。在后端优化模块中，引入神经辐射场(NeRF)几何增强模块，为每个特征点提供深度先验，协助优化器在弱纹理与动态场景中恢复更加准确和稳定的三维结构估计。通过公开数据集EuRoc测试，实验结果表明系统在改进算法过后在更多复杂环境会有更好的定位、建图和实时性方面表现。

Abstract: Aiming at the requirements of autonomous positioning and navigation of UAVs in complex environments (such as indoor, night, and dynamic scenes) where GPS signals are lost or unavailable, an optimization method for high-performance visual inertial odometer (VIO) system is proposed and implemented. The system integrates the data of the vision sensor and the inertial measurement unit (IMU) to obtain the final UAV pose information, but the traditional visual inertial odometer method is susceptible to weak texture, dynamic lighting, non-Gaussian noise and appearance changes. The method uses CLAH local adaptive histogram equalization and Gamma correction algorithm to enhance the image at the front end of the image, which effectively improves the image quality in low-light and low-texture scenes. At the same time, the GFTT BRISK combined feature extraction strategy is introduced to improve feature stability and matching robustness. In the back-end optimization module, the Neural Radiation Field (NeRF) geometry enhancement module is introduced to provide depth priors for each feature point, helping the optimizer to restore more accurate and stable 3D structure estimation in weak textures and dynamic scenes. Through the public dataset EuRoc test, the experimental results show that the system will perform better in positioning, mapping and real-time performance in more complex environments after the improved algorithm.