Vehicle tracking in single camera systems

master's thesis

• Title (croatian): Praćenje vozila u sustavima s jednom kamerom
• Author: Florijan Iljazović
• Mentor: Bojan Basrak (mentor)
• Committee member: Bojan Basrak (predsjednik povjerenstva)
• Committee member: Zoran Vondraček (član povjerenstva)
• Committee member: Siniša Slijepčević (član povjerenstva)
• Committee member: Snježana Lubura Strunjak (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Mathematics); Zagreb
• Defense date and country: 2020-09-30, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Mathematics

Abstract

Neuralne mreže i duboko učenje su moćan i brzo rastući alat prilikom analize slika prometa, čemu pomažu brojni javno dostupni resursi za treniranje i testiranje neuronskih mreža na velikim skupovima podataka kao što je KITTI. To je moguće iskoristiti i povratne podatke iz CNN sistema za raspoznavanje objekata (kamera je montirana na vozilu) poslati u odgovarajući algoritam za praćenje u realnom vremenu. Ovaj rad koristi recentnu metodu ([9]) dobivanja detekcija vozila prisutnih na slici i Kalmanov filter kao algoritam praćenja. Testiranje se provodi na KITTI benchmarku. U prvom poglavlju navodimo neka bitna svojstva višedimenzionalne normalne distribucije, s naglaskom na Bayesov teorem za normalnu razdiobu, koji se koristi u izvodu Kalmanovog filtra. U drugom poglavlju, nakon kratkog uvoda u područje algoritama za praćenje I odgovarajuću terminologiju, koristimo bayesovski pristup kako bismo dobili općenito rekurzivno rješenje za praćenje jednog objekta. Zatim diskutiramo i detaljno izvodimo Kalmanov filter, a kraće komentiramo i njegove bitne ekstenzije, prošireni Kalmanov filter i unscented Kalmanov filter. U trećem poglavlju dajemo kratki osvrt na moderne metode raspoznavanja vozila i njihovo praćenje, te prezentiramo našu metodu. U četvrtom poglavlju navodimo neke implementacijske detalje i iznosimo rezultate na KITTI podacima. Glavni problem sustava u odnosu na ekvivalent baziran na lidaru je manji broj deketiranih objekata. Usprkos tome, rezultati su solidni.

Abstract (english)

Neural networks and deep learning are a powerful tool for detection of vehicles in camera images, growing steadily due to advance in ideas and large-scale public benchmarks such as KITTI. In principle, this can be exploited by using output of appropriate deep CNN system and feeding it as measurements input to the object tracking algorithm. This work combines state of the art monocular detection system ([9]) with the Kalman filter algorithm into multiple object tracking pipeline. The performance is tested on public KITTI benchmark video sequences. In the first chapter we review some important properties of multivariate normal distribution, with the stress on the Bayes' theorem for Gaussian densities, which is empoyed in the derivation of the Kalman fillter in the subsequent chapter. In chapter 2, after giving some introduction to object tracking problem and terminology, we first consider more general tracking setup employing Bayesian probabilistic framework. Then we give detailed derivation of Kalman fiter, and overview of its notable variations, extended Kalman filter and unscented Kalman filter. In chapter 3 we shortly review state of the art object detection and tracking systems, and present our method. In chapter 4 we remark some implementation details and report results of our method on the KITTI tracking benchmark. While it achieves solid performance, detector has too many false negatives (missed target) to achieve state of the art results

• Keywords: neuralne mreže; duboko učenje; KITTI; Kalmanov filter
• Keywords (english): neural networks; deep learning; KITTI; Kalman fillter
• Language: croatian
• URN:NBN: urn:nbn:hr:217:552541
• Study programme: Title: Mathematical Statistics; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra matematike (magistar/magistra matematike)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2021-02-25 12:59:39