Vis enkel innførsel

dc.contributor.advisorHa, Hoai-Phuong
dc.contributor.authorPhan, Hung Ngoc
dc.date.accessioned2022-05-31T12:06:01Z
dc.date.available2022-05-31T12:06:01Z
dc.date.issued2021-05-31
dc.description.abstractBackground modeling is a promising field of study in video analysis, with a wide range of applications in video surveillance. Deep neural networks have proliferated in recent years as a result of effective learning-based approaches to motion analysis. However, these strategies only provide a partial description of the observed scenes' insufficient properties since they use a single-valued mapping to estimate the target background's temporal conditional averages. On the other hand, statistical learning in the imagery domain has become one of the most widely used approaches due to its high adaptability to dynamic context transformation, especially Gaussian Mixture Models. Specifically, these probabilistic models aim to adjust latent parameters to gain high expectation of realistically observed data; however, this approach only concentrates on contextual dynamics in short-term analysis. In a prolonged investigation, it is challenging so that statistical methods cannot reserve the generalization of long-term variation of image data. Balancing the trade-off between traditional machine learning models and deep neural networks requires an integrated approach to ensure accuracy in conception while maintaining a high speed of execution. In this research, we present a novel two-stage approach for detecting changes using two convolutional neural networks in this work. The first architecture is based on unsupervised Gaussian mixtures statistical learning, which is used to classify the salient features of scenes. The second one implements a light-weighted pipeline of foreground detection. Our two-stage system has a total of approximately 3.5K parameters but still converges quickly to complex motion patterns. Our experiments on publicly accessible datasets demonstrate that our proposed networks are not only capable of generalizing regions of moving objects with promising results in unseen scenarios, but also competitive in terms of performance quality and effectiveness foreground segmentation. Apart from modeling the data's underlying generator as a non-convex optimization problem, we briefly examine the communication cost associated with the network training by using a distributed scheme of data-parallelism to simulate a stochastic gradient descent algorithm with communication avoidance for parallel machine learningen_US
dc.identifier.urihttps://hdl.handle.net/10037/25327
dc.language.isoengen_US
dc.publisherUiT Norges arktiske universiteten_US
dc.publisherUiT The Arctic University of Norwayen_US
dc.rights.accessRightsopenAccessen_US
dc.rights.holderCopyright 2021 The Author(s)
dc.rights.urihttps://creativecommons.org/licenses/by-nc-sa/4.0en_US
dc.rightsAttribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)en_US
dc.subject.courseIDINF-3981
dc.subjectVDP::Mathematics and natural science: 400::Information and communication science: 420::Simulation, visualization, signal processing, image processing: 429en_US
dc.subjectVDP::Matematikk og Naturvitenskap: 400::Informasjons- og kommunikasjonsvitenskap: 420::Simulering, visualisering, signalbehandling, bildeanalyse: 429en_US
dc.subjectVDP::Mathematics and natural science: 400::Information and communication science: 420::Communication and distributed systems: 423en_US
dc.subjectVDP::Matematikk og Naturvitenskap: 400::Informasjons- og kommunikasjonsvitenskap: 420::Kommunikasjon og distribuerte systemer: 423en_US
dc.titleInvestigating the latency cost of statistical learning of a Gaussian mixture simulating on a convolutional density network with adaptive batch size technique for background modelingen_US
dc.typeMaster thesisen_US
dc.typeMastergradsoppgaveen_US


Tilhørende fil(er)

Thumbnail
Thumbnail
Thumbnail

Denne innførselen finnes i følgende samling(er)

Vis enkel innførsel

Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)
Med mindre det står noe annet, er denne innførselens lisens beskrevet som Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)