Bultmann, Simon Alexander: Semantic Feedback for Collaborative Perception with Smart Edge Sensors. - Bonn, 2024. - Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5-76533
@phdthesis{handle:20.500.11811/11943,
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5-76533,
author = {{Simon Alexander Bultmann}},
title = {Semantic Feedback for Collaborative Perception with Smart Edge Sensors},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2024,
month = aug,

note = {In this thesis, we develop a system for accurate semantic perception of 3D scene geometry, persons, and objects in robotic applications. We consider the limitations of interpreting only a single sensor view with restricted measurement range, field of view, and resolution, and the challenges posed by centralized approaches that rely on high communication bandwidth and computational power.
To address these issues, we propose a network of distributed smart edge sensors equipped with a multi-modal sensor suite and an embedded CNN inference accelerator for on-device image processing. Real-time vision CNN models for person and object detection, semantic segmentation, and pose estimation are deployed on the sensors. The extracted information, such as 2D human keypoints, object poses, and semantic point clouds, is then passed to a central backend where multiple viewpoints are fused into a comprehensive 3D semantic scene model. Since image interpretation is computed locally, only semantic information is sent over the network. The raw images remain on the sensor boards, significantly reducing bandwidth requirements and mitigating privacy concerns for the observed persons.
The concept of smart edge sensors is further extended to mobile aerial and ground robots, enabling anticipatory human-aware navigation and active perception in areas not covered by stationary sensors. An outdoor smart edge sensor is presented, based on a UAV platform with on-board multi-modal semantic perception.
We introduce the concept of semantic feedback, enabling collaborative perception between sensor nodes and the central backend through bidirectional communication at the semantic level. The incorporation of global context information, such as the fused multi-view human and robot pose estimates, enhances the local semantic models of the smart edge sensors, improving pose estimation and enabling preemptive adjustments to the robot's navigation path, e.g. when a person emerges from an occluded area.
The proposed methods are evaluated using public datasets and real-world experiments in challenging cluttered and dynamic environments. The system demonstrates the ability to generate a real-time semantic scene model that includes semantically annotated 3D geometry, object instances, and poses of multiple persons.},

url = {https://hdl.handle.net/20.500.11811/11943}
}

The following license files are associated with this item:

InCopyright