Explorative Comparison Between Classic Computer Vision Techniques and Deep Learning: Comparing performance of computer vision approaches in dynamic outdoor sceneries
Translated title
Explorative Comparison Between Classic Computer Vision Techniques and Deep Learning
Author
Pedersen, Kristian Hyttel
Term
4. semester
Education
Publication year
2021
Submitted on
2021-06-03
Pages
33
Abstract
To make agricultural vehicles drive autonomously, one practical approach is to follow the permanent lanes that tractors leave in fields. These lanes are almost always present and fairly stable, and keeping vehicles to them typically increases crop yield by 5–15%. This project uses a standard color (RGB) camera to detect those lanes and compares two methods: a hand-crafted “classic” computer vision pipeline built from scratch and a neural network based on YOLOv3, a popular model for object detection. Both approaches detect lanes reasonably well. The classic system almost always produces a usable estimate to guide a vehicle, but it is more prone to incorrect outputs than the neural network. In contrast, the neural network often makes no prediction in some test images, yet it almost never makes false detections. The study concludes that a classic approach is well suited for short timelines or proof-of-concept work, while a neural network, given more training or a larger dataset, is better for large-scale projects due to higher accuracy. Although the use case is niche, the same principles may transfer to similar tasks in other domains.
For at gøre landbrugskøretøjer selvkørende kan de følge de faste kørespor, som traktorer efterlader på marken. Disse spor er næsten altid til stede og forholdsvis permanente, og alene at holde kørsel til faste spor øger typisk udbyttet med 5–15 %. I denne opgave bruges et almindeligt farvekamera (RGB) til at finde sporene, og to metoder sammenlignes: et klassisk computer-vision-system bygget fra bunden (en håndlavet billedbehandlingsløsning) og et neuralt netværk baseret på YOLOv3, en udbredt model til objektdetektion. Begge tilgange kan finde kørespor med rimelig succes. Det klassiske system giver næsten altid et resultat, der kan bruges til at styre et køretøj, men er mere tilbøjeligt til forkerte resultater end det neurale netværk. Omvendt forudsiger det neurale netværk ofte ingenting i nogle testbilleder, men laver næsten aldrig falske forudsigelser. Konklusionen er, at et klassisk system er velegnet ved korte tidsfrister eller som proof-of-concept, mens et neuralt netværk med mere træning eller et større datasæt er bedre til storskalaprojekter på grund af mere præcise resultater. Selvom use casen er snæver, kan principperne overføres til lignende opgaver i andre domæner.
[This apstract has been rewritten with the help of AI based on the project's original abstract]
Keywords
Computer Vision ; Neural Network ; AI ; Agriculture ; Autonomous Vehicle ; Autonomous Vehicles ; Autonomous ; OpenCV ; YOLO ; YOLOv3