Vairākuzdevumu mācīšanās pieejas analīze datorredzes uzdevumos
Loading...
Date
Authors
Advisor
Journal Title
Journal ISSN
Volume Title
Publisher
Latvijas Universitāte
Language
lav
Abstract
Mūsdienu transporta sistēmas arvien vairāk balstās uz automatizētām datorredzes tehnoloģijām, kas spēj pastāvīgi analizēt satiksmes datus. Šajā kontekstā aktuāla ir vairākuzdevumu mācīšanās (MTL) pieeja, kura sniedz potenciālus uzlabojumus precizitātē un resursu patēriņā. Darba mērķis ir veikt MTL pieejas analīzi, pētot tās parametrus (uzdevumu veids, svaru koeficienti un optimizēšanas stratēģijas) un salīdzināt to ar individuālo uzdevumu mācīšanās pieeju. Darbā tika izstrādāti modeļi transportlīdzekļa lokalizēšanai un tā horizontālā virziena noteikšanai, izmantojot klasifikācijas un regresijas uzdevumus. Rezultāti liecina, ka MTL pieeja nodrošina priekšrocības resursu patēriņā, taču kopējais pieejas sniegums ir atkarīgs no izmantotās svaru balansēšanas stratēģijas, kur nepareizi izvēlēti svaru koeficienti rezultātus var pasliktināt.
Intelligent transportation systems increasingly rely on automated computer vision technologies that can independently analyse traffic data. In this context, the multi-task learning (MTL) is relevant, because it offers potential improvements in accuracy and resource consumption. The aim of this work is to analyse MTL approach by studying its parameters (task type, weight coefficients, and optimization strategies) and comparing it with single task learning approach. Within the study, models were developed for vehicle localization and horizontal direction estimation as classification and regression tasks. The results indicate that MTL approach provides advantages in resource consumption, however, the overall performance depends on the used weight balancing strategy, where incorrectly chosen weight coefficients can worsen the results.
Intelligent transportation systems increasingly rely on automated computer vision technologies that can independently analyse traffic data. In this context, the multi-task learning (MTL) is relevant, because it offers potential improvements in accuracy and resource consumption. The aim of this work is to analyse MTL approach by studying its parameters (task type, weight coefficients, and optimization strategies) and comparing it with single task learning approach. Within the study, models were developed for vehicle localization and horizontal direction estimation as classification and regression tasks. The results indicate that MTL approach provides advantages in resource consumption, however, the overall performance depends on the used weight balancing strategy, where incorrectly chosen weight coefficients can worsen the results.