一种结合深度学习的运动去模糊视觉SLAM方法

针对高动态环境下视觉同步定位与地图构建（Simultaneous Localization And Mapping，SLAM）系统的可靠性受运动模糊的限制，研究了一种基于生成对抗网络(Generative Adversarial Network，GAN)和AKAZE特征点的运动去模糊视觉SLAM方案。首先对因相机快速运动而产生的模糊图像进行AKAZE特征点的提取与检测，并根据特征点分布的丰富程度计算图像块权重，结合灰度图像的方差信息建立特征点与模糊程度之间的量化关系表；之后将达到模糊分数阈值的图像同步输入至改进GAN网络模型，该网络以端对端的形式恢复中心模糊帧的纹理信息，最后将输出的清晰图像重新进行位姿估计参与ORB-SLAM2后端优化过程。在公开数据集TUM上进行测试，对于受模糊影响较严重的数据集，方案可以明显降低相机轨迹估计的整体误差，同时维持较好的鲁棒性。

A Motion Deblurring Visual SLAM Method Combined With Deep Learning

Aiming at the limitation of motion blur on the reliability of visual simultaneous localization and mapping system in high dynamic environment, a motion deblurring visual SLAM scheme based on generative adversarial network and AKAZE feature points is studied. Firstly, the AKAZE feature points are extracted and detected for the blurred image generated by the rapid movement of the camera, and the image block weight is calculated according to the richness of feature points distribution. Combined with the variance information of gray image, the quantitative relationship table between feature points and blur degree is established; After that, the image reaching the fuzzy score threshold is synchronously inputted to the improved GAN network model. The network restores the texture information of the central fuzzy frame in the form of end-to-end. Finally, the output clear image is re pose estimated to participate in the ORB-SLAM2 back-end optimization process. Tested on the open data set tum, for the data set seriously affected by ambiguity, the scheme can significantly reduce the overall error of camera trajectory estimation and maintain good robustness.