Analiza uspešnosti algoritmov za zaznavanje objektov, naučenih na zamegljenih slikah

Planinšek, David (author), Perš, Janez (mentor), Ivanovska, Marija (comentor)

Namen te diplomske naloge je preizkusiti algoritme za videonadzor prometa ob ohranjanju visoke ravni zasebnosti podatkov. Deidentifikacija podatkov je bila dosežena z zameglitvijo slik, ki so bile nato uporabljene za učenje in testiranje algoritmov za detekcijo vozil. Za učenje modelov smo uporabili slike, ki predstavljajo različne stopnje zamegljenosti. Vsakič smo za učenje izbrali modele slik z manjšo stopnjo zamegljenosti, nato pa smo naučene algoritme preizkusili na slikah z večjo ali enako stopnjo zamegljenosti. V okviru diplomske naloge smo uporabili štiri različne algoritme za prepoznavanje objektov na slikah: DETR, Faster R-CNN, YOLOv3 in HRNet. DETR (ang. Detection Transformer) je nov pristop, ki uporablja transformerje za neposredno prepoznavanje objektov na sliki brez potrebe po predhodnih predmetnih predlogah. Faster R-CNN je bil izbran zaradi njegove hitrosti in natančnosti pri prepoznavanju objektov, YOLOv3 (ang. You Only Look Once) je znan po svoji učinkovitosti pri realnočasovni detekciji objektov, HRNet (ang. High-Resolution Network) pa je znan po svoji natančnosti pri segmentaciji objektov, zlasti pri slikah z visoko ločljivostjo. Baza označenih slik, uporabljena v nalogi, je sestavljena iz več tisoč slik, ki vključujejo avtomobile, avtobuse in kombije. Slike so bile ročno označene z uporabo očrtanih pravokotnikov (ang. bounding boxes), ki določajo položaj in velikost objektov na sliki. Slike v bazi so bile razdeljene na tri kategorije glede na stopnjo zamegljenosti: jasne slike, prva stopnja zamegljenosti in druga stopnja zamegljenosti slik. To omogoča sistematično preučevanje učinkovitosti algoritmov pri različnih stopnjah zamegljenosti slik. Končne rezultate smo delili na kvalitativno in kvantitativno analizo. Kvantitativna analiza je vključevala pregled slik in ročno analizo rezultatov, medtem ko smo za kvalitativno analizo uporabili program, ki je izvedel analitično računanje uspešnosti prepoznavanja objektov na slikah. S tem smo dobili celovit vpogled v uspešnost in zmogljivosti vsakega algoritma pri različnih stopnjah zamegljenosti slik.

• Language: Slovenian
• Keywords: prepoznavanje objektov, zamegljene slike, cestni promet, vgrajena zasebnost, algoritmi, baza označenih slik
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FE - Faculty of Electrical Engineering
• Year: 2024
• PID: 20.500.12556/RUL-159390
• COBISS.SI-ID: 201174275
• Publication date in RUL: 09.07.2024

PLANINŠEK, David, 2024, Analiza uspešnosti algoritmov za zaznavanje objektov, naučenih na zamegljenih slikah [online]. Bachelor’s thesis. [Accessed 23 April 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=159390

Secondary language

• Language: English
• Title: Analysis of Object Detection Algorithms Trained on Blurred Images

Abstract

This thesis focuses on employing commonly used object detection algorithms for traffic video monitoring under privacy-preserved conditions. The deidentification of the data was achieved by mechanically defocusing the camera used for image acquisition. Blurred images and their sharp counterparts were then used to train and test the efficiency of different object detection methods. Within the scope of this thesis, four distinct algorithms were employed for object detection in images: DETR (ang. Detection Transformer), Faster R-CNN, YOLOv3, and HRNet. DETR introduces a novel approach that utilizes transformers for direct object detection in images without needing prior object proposals. Faster R-CNN was chosen for its speed and accuracy in object detection. YOLOv3 is known for its efficiency in real-time object detection, and HRNet is recognized for its precision in object segmentation, particularly in high-resolution images. The annotated image database used in this research includes thousands of images featuring cars, buses, and vans, all manually annotated with precise bounding boxes to define the position and size of objects within the images. The database categorizes these images into three blurriness levels: clear, moderately blurred, and heavily blurred. This categorization enables a systematic examination of algorithm performance across different levels of image blurriness. The thesis concludes with both qualitative and quantitative analyses. The quantitative analysis involves reviewing images and manually analyzing the outcomes, whereas a software program performs the qualitative analysis, executing analytical computations of object detection efficacy. This comprehensive evaluation provides insights into the performance and capabilities of each algorithm under various levels of image blurriness.

• Keywords: object detection, blurred images, road traffic, embedded privacy, algorithms, annotated image database