Utilizing semantics for fast and robust localisation and mapping in semi-structured environments

Zaganidis, Anestis (2019) Utilizing semantics for fast and robust localisation and mapping in semi-structured environments. PhD thesis, University of Lincoln.

Thesis submission form
Thesis submission form
Utilizing semantics for fast and robust localisation and mapping in semi-structured environments
PhD Thesis
ZAG15614958 Research Electronic Submission.pdf - Supplemental Material

Zaganidis, Anestis - Computer Science - April 2020.pdf - Whole Document

Item Type:Thesis (PhD)
Item Status:Live Archive


Autonomous Navigation in outdoor, semi structured environments such as
agricultural fields, warehouses and public transport is crucial for the seamless
operation of autonomous vehicles. Simultaneous localization and mapping
(SLAM), the online construction of a map and the localization within it,
remains a challenging problem, especially for these outdoor environments
with low structure, where the pose has 6 degrees of freedom. This thesis
addresses the problem of localisation and mapping in three dimensional, semi structured environments using a lidar range sensor and proposes a novel
SLAM system assisted by semantic segmentation of point clouds.
This work makes extensive use of the Normal Distributions Transform
(NDT), a compact representation that can be used for point cloud registration and place recognition, and proposes extensions that increase the robustness and speed in semi-structured environments. In the absence of strong
geometric features, semantics can decrease the rate of incorrect registration
correspondences in point cloud registration, and can increase the specificity
of the NDT Histogram descriptor for place recognition.
The main contributions of this work include, (i) a comprehensive review
of the registration and loop closure detection algorithms (ii) the integration of
semantics into the Normal Distributions Transform registration using handcrafted features as well as data-driven semantic segmentation models (iii) the
reuse of the computed semantics for loop closure detection.
In this work, we study two methods for the extraction of semantics. Handcrafted features provide an input for low-level semantics that improve registration accuracy, robustness and speed in semi-structured scenes with high overlap. A data-driven classifier is used to provide high-level, human interpretable semantics, that when used for registration increases robustness to
initial registration errors to similar levels as global registration methods. The same data-driven classifier is also used for the loop closure detection module of the SLAM pipeline. This classifier, along with the inclusion of semantics in the NDT Histograms descriptor and engineered rules, is capable of operating with zero false positives in a semi-structured environment. In contrast,
the non-semantic version exhibits high rate of false-positives that is highly limiting for outdoor SLAM applications. All the proposed components of the system are extensively evaluated with the use of publicly available datasets. The complete pipeline, including pretrained models, is released as an open-source ROS package.

Keywords:semantics, localisation, mapping, semi-structured environments
Subjects:G Mathematical and Computer Sciences > G400 Computer Science
Divisions:College of Science > School of Computer Science
ID Code:49976
Deposited On:28 Jun 2022 13:49

Repository Staff Only: item control page