Automated guided vehicle localization: A mapping strategy in dynamic environments, using computer vision

Abstract

Many warehouses rely on automated guided vehicles (AGVs) that use 2D-LiDAR based localization and static environmental landmarks. In dynamic warehouse environments, this approach can become unreliable due to occlusions and changes of previously considered static landmarks. A vision based localization method is able to detect features in the environment that a 2D-LiDAR cannot. Therefore, this thesis aims to answer the main question: "Can a visual localization method improve localization accuracy compared to LiDAR-based 2D-localization, based on common accuracy metrics like absolute trajectory error (ATE)?" To evaluate our own solution, a baseline was created consisting of a modified version of the open-source method pySLAM [1]. The baseline was modified to rely on a static map for localization, reflecting how many industrial navigation systems operate. Based on the observed limitations of this approach in dynamic environments, a new method called 2P (two point)-SLAM was developed and made available as open source [2]. The method separates map points into static and mutable points, allowing the system to preserve a pre-recorded static map while adapting to environmental changes by adding and removing mutable points when needed. The results show that the maximum error of the baseline is 307 mm during the dynamic test. SLAM achieves a maximum error of 144 mm, while 2P-SLAM achieves 184 mm. However, 2P-SLAM adds 83% fewer points than SLAM. The ATERMSE suggests that there is no significant difference between 2P-SLAM and the baseline, with values of 76 mm and 78 mm respectively. As expected, the existing 2D-LiDAR solution struggles in the dynamic environment with an ATERMSE of 309 mm. The conducted experiments therefore suggest that a vision based localization method can improve localization accuracy compared to the evaluated 2D-LiDAR localization method in the tested dynamic environment. Since each localization method was evaluated using a single execution and the visual localization pipeline exhibits stochastic behaviour, the reported results should be interpreted as indicative rather than statistically significant. Further testing and repeated evaluations are required to strengthen these findings.