AAU Student Projects is unavailable between June 15th 1.30pm and 17th 1.30pm due to planned system maintenance. The projects cannot be downloaded during this period.
AAU Student Projects - visit Aalborg University's student projects portal
An executive master's programme thesis from Aalborg University
Book cover


Deep Learning Based Hazard Detection for Bike Paths

Authors

; ;

Term

4. term

Publication year

2026

Pages

58

Abstract

Bike-path hazards, such as cracks, potholes, barriers, and debris, pose safety risks and need effective monitoring. This study presents a YOLOv8-based machine learning method to automatically detect these hazards. We collected 75% of the data in Aalborg, Denmark, using a smartphone mounted on a bicycle handlebar in varied environment, weather, and lighting conditions; the remaining data came from the internet. The dataset was then refined and optimized to improve data quality and model performance. Because small hazards are especially important, particularly cracks, we used image tiling, which splits images into smaller patches, to improve the detection and localization of small objects. The model achieved a precision of 0.801 and an mAP0.5 of 0.786. Compared with training on the untiled dataset, precision and mAP0.5 increased by 29.6% and 47.6%. Precision is the share of detections that are correct, and mAP0.5 is a standard accuracy metric. These results support the effectiveness of context-aware data collection combined with image tiling for improving small-object detection.

Farlige forhold på cykelstier, såsom revner, huller, forhindringer og affald, udgør en sikkerhedsrisiko og kræver effektiv overvågning. Dette studie præsenterer en maskinlæringsmetode baseret på YOLOv8 til automatisk at opdage sådanne farer. 75% af data blev indsamlet i Aalborg, Danmark, med en mobiltelefon monteret på cykelstyret under forskellige miljø-, vejr- og lysforhold; resten blev hentet fra internettet. Det indsamlede datasæt blev efterfølgende renset og optimeret for at forbedre datakvalitet og modelpræstation. Vi fokuserer særligt på små farer, især revner. For bedre at opdage og lokalisere små objekter brugte vi billedopdeling (image tiling), hvor billeder deles op i mindre felter. Modellen opnåede en precision på 0.801 og en mAP0.5 på 0.786. Sammenlignet med en model trænet på et ikke-opdelt datasæt steg precision og mAP0.5 med henholdsvis 29.6% og 47.6%. Precision angiver andelen af korrekte detektioner, og mAP0.5 er en standardmåling af nøjagtighed. Resultaterne understøtter, at kontekstbevidst dataindsamling kombineret med billedopdeling forbedrer detektion af små objekter.

[This apstract has been rewritten with the help of AI based on the project's original abstract]