关键词: GNSS随机模型, 鱼眼相机, 神经网络, 注意力机制, 城市复杂场景

Abstract:

Global navigation satellite system (GNSS) can provide high-precision positioning services. However, in complex urban environments, the presence of multipath effects and non-line-of-sight (NLOS) signals leads to a mismatch between GNSS observation quality and prior stochastic models, significantly degrading positioning performance. Methods based on fisheye cameras can utilize sky-view information to mitigate the impact of NLOS observations, but existing solutions are mostly limited to semantic segmentation applications and fail to fully exploit the high-dimensional environmental features in images. To address this issue, this paper proposes a neural network-based GNSS stochastic model generation method using fisheye images. The proposed method employs neural networks to extract high-dimensional environmental features from images that reflect GNSS observation conditions and tightly integrates GNSS and image features in a cross-attention layer to predict the stochastic model of satellite observations. Experimental results demonstrate that the proposed method can effectively capture the correlation between fisheye images and GNSS observation environments, accurately inflating the variance of abnormal observations. Moreover, in scenarios where fisheye images are affected by errors, the method can leverage GNSS feature information to reduce the impact of image errors on prediction results. When further applied to a RTK/IMU integrated navigation system, the proposed method improves positioning accuracy by 32.9%, verifying that the proposed method can significantly reduce the influence of abnormal observations and enhance system performance in complex urban environments.

Key words: GNSS stochastic model, fisheye camera, neural networks, attention mechanism, complex urban scenes