通过深度强化学习(DRL)生成的大规模2D SLAM错误数据集
收藏资源简介:
本文提出了一种名为TT-OGM的新方法,该方法通过深度强化学习(DRL)生成大规模2D SLAM错误数据集,用于训练生成对抗网络(GAN),以改善SLAM算法在复杂场景中的定位和建图质量。TT-OGM方法通过转换和翻译步骤,将3D点云数据转换为2D表示,从而在具有动态运动的复杂场景中实现高精度和可靠的2D OGM。该方法还集成了一个基于GAN的观测完成和错误删除模块,能够检测和删除传感器噪声、无关数据和伪影,同时完成部分映射区域,部分观测和重新对齐角/线性偏移。
This paper proposes a novel method named TT-OGM, which generates large-scale 2D SLAM error datasets via deep reinforcement learning (DRL) for training generative adversarial networks (GANs) to enhance the localization and mapping quality of SLAM algorithms in complex scenarios. The TT-OGM method converts 3D point cloud data into 2D representations through transformation and translation steps, thereby achieving high-precision and robust 2D OGM in complex scenes with dynamic motions. Furthermore, this method integrates a GAN-based observation completion and error removal module that can detect and eliminate sensor noise, irrelevant data and artifacts, while completing partially mapped regions, performing partial observation and realigning angular/linear offsets.




