AAU Student Projects - visit Aalborg University's student projects portal
A master's thesis from Aalborg University
Book cover


Identification of Cigarette Litter with the use of Outdoor Mobile Robots

Author

Term

4. semester

Education

Publication year

2021

Submitted on

Pages

96

Abstract

Cigaretskod er udbredt, skadeligt for miljøet og dyrt at fjerne manuelt. I samarbejde med SkodRobot-projektet (Aarhus Kommune, Capra Robotics og KK-tech) adresserer denne afhandling den første del af et todelt mål: at gøre en udendørs mobil robot i stand til at identificere cigaretaffald på fortove i realtid som grundlag for efterfølgende autonom fjernelse. Arbejdet omfatter opbygning af et specialiseret datasæt for cigaretskod med annotering og dataaugmentering, valg og træning af en Tiny-YOLOv4-baseret detektor i Darknet, integration i ROS via Darknet-ROS og implementering på en Nvidia Jetson Nano med afprøvet CSI-kameraopsætning. Derudover undersøges metoder til lokalisering via monokulær afstandsestimering for at koble detektioner til robotens handlinger. Systemet monteres på Capra Hircus, en udendørs mobilplatform, og afprøves på billeder, video og under kørsel for at vurdere robusthed og kravopfyldelse. Afhandlingen rapporterer den nuværende fremdrift: et funktionsdygtigt system er udviklet, integreret og testet på robotten; detaljerede kvantitative resultater fremgår ikke her, og fremtidigt arbejde omfatter udvidelse af datasættet samt forbedring af detektion og autonom dækning.

Cigarette-butt litter is widespread, environmentally harmful, and costly to remove manually. In collaboration with the SkodRobot project (Aarhus Municipality, Capra Robotics, and KK-tech), this thesis addresses the first stage of a two-part goal: enabling an outdoor mobile robot to identify cigarette litter on sidewalks in real time as a foundation for subsequent autonomous removal. The work builds a dedicated cigarette-litter dataset with annotation and data augmentation, selects and trains a Tiny-YOLOv4-based detector using the Darknet framework, integrates it with ROS via Darknet-ROS, and deploys it on an Nvidia Jetson Nano with a tested CSI camera setup. Methods for localization through monocular distance estimation are explored to link detections to robot actions. The complete system is mounted on the Capra Hircus outdoor platform and tested on images, videos, and during robot runs to assess robustness and compliance with requirements. The thesis reports current progress: a working system has been developed, integrated, and tested on the robot; detailed quantitative metrics are not provided here, and future work focuses on expanding the dataset and improving detection and autonomous coverage.

[This summary has been generated with the help of AI directly from the project (PDF)]