Visual Navigation Technology of Monocular Fisheye Camera Based on Reinforcement Learning

Current methods have problems such as missing scale information, severe distortion, and poor adaptability in unmanned visual imaging systems. This paper studies the visual navigation technology of monocular fisheye cameras based on multi-task learning, integrating multiple functions such as depth estimation, distortion correction, visual odometer, semantic segmentation, and target detection. By analyzing the characteristics of each task, a complete image processing system is finally formed. This paper adopts a depth estimation algorithm based on self-supervised learning, and uses deep learning methods to correct image distortion, improve image quality and navigation accuracy, and adopts a joint training strategy to optimize the model's adaptability to dynamic environments. The multi-task learning system has an SSIM value of up to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 9 5}$</tex> in the scale information loss evaluation, a feature point matching accuracy of up to 97.1 %, and the root mean square error of relative pose in different dynamic scenes is the smallest, showing superior performance and robustness.