| Sažetak | Poznavanje mobilnosti korisnika prometne mreže osnova je za određivanje prometne potražnje i planiranje transporta. Podatci s mobilnih uređaja postali su vrijedan izvor prometnih podataka o mobilnosti korisnika u prometnoj mreži. Za prikupljanje dodatnih informacija tijekom kretanja korisnika prometnom mrežom, poput modova prijevoza, koriste se razne metode kojima se iz prikupljenih podataka određuju informacije koje se ne mogu automatski prikupljati. Za detekciju moda prijevoza koriste se metode klasifikacije trajektorije, pri čemu se klasificira dio trajektorije u kojem je korišten jedan mod prijevoza. Trajektoriju je prije klasifikacije potrebno podijeliti na segmente koji sadrže jedan mod prijevoza. U literaturi su zastupljene mnogobrojne metode za klasifikaciju modova prijevoza koje se najčešće zasnivaju na metodama strojnog učenja dok su za segmentaciju trajektorije zastupljene metode koje se zasnivaju na nizu iskustveno postavljenih pravila. Nedostatak tako generiranih pravila je što su često prilagođena skupu podataka nad kojim su razvijena i nisu jedinstveno primjenjiva s jednakom točnošću na druge skupove podataka. U ovoj doktorskoj disertaciji je razvijena višerazinska metoda za segmentaciju trajektorije i klasifikaciju modova prijevoza u stvarnom vremenu. Za odlučivanje se koriste matrice prijelaznih stanja i algoritam slučajnih šuma. Metoda je razvijena i testirana koristeći objavljene skupove podataka, a za validaciju je korišten novi skup podataka prikupljen u sklopu ovog istraživanja. U prvom dijelu doktorske disertacije detaljno je opisana implementacija višerazinske metode za segmentaciju i klasifikaciju trajektorije. Također, opisan je postupak obrade podataka koji uključuje podjelu sirovih podataka u vremenske prozore te preslikavanje podataka iz vremenske u frekvencijsku domenu. Na kraju prvog dijela doktorske disertacije analizirani su rezultati klasifikacije trajektorije razvijenom metodom i uspoređeni su s rezultatima postignutim na testnom skupu podataka dostupnom u literaturi. U drugom dijelu doktorske disertacije opisan je novi skup podataka prikupljen u sklopu ovog istraživanja s naglaskom na mobilnu aplikaciju i način prikupljanja podataka te distribuciju modova prijevoza u skupu podataka. Prikupljeni skup podataka korišten je za validaciju metode čiji su rezultati prikazani na kraju ovog dijela. |
| Sažetak (engleski) | This doctoral dissertation is organized into five chapters, and the first three provide the introduction and motivation for developing a multilevel method for real-time trajectory segmentation and classification. The fourth chapter describes the newly developed multilevel method for trajectory classification and segmentation, and presents the results achieved on the publicly available test dataset. The fifth chapter describes the developed mobile application for collecting data and newly collected dataset. In the same chapter, validation results for the multilevel method on the newly collected dataset are presented. Finally, the conclusion of the paper and topics for future work are given in the last chapter. The first introductory chapter gives an overview of the area of urban mobility to which the problem of trajectory segmentation and classification according to transport mode belongs. It provides the motivation for this research and an overview of the main scientific contributions in this research. The second chapter describes the problem of transport mode classification, emphasizing the classification of segmented and non-segmented trajectories. The most frequently used data are described, and an overview of public and private datasets used in the literature is given. Also, relevant features used in the literature to classify transport modes are presented. The emphasis of the third chapter is on methods for solving the problem of transport mode classification. Basic trajectory segmentation methods from the supervised, unsupervised and semi-supervised methods are described. Also, an overview of data-driven methods for transport mode classification is given. At the end of the chapter, the structure of the test problem is described based on the Sussex-Huawei Locomotion (SHL) dataset used in this research for method training and testing. Accuracy is measured by the recall and precision ratio and the overall accuracy of the method. The fourth chapter provides a detailed description of the multilevel method for real-time trajectory segmentation and classification. The method includes four basic components: (i) data preprocessing, (ii) method for trajectory segmentation, (iii) transport mode classification method, and (iv) rules in the multilevel method. During data preprocessing, the data are divided into time windows to simulate a real-time environment, and the features that are the input data for the multilevel method are calculated. In order to discover information from raw data, the discrete Fourier transform procedure was used to transform data from the time domain to the frequency domain. Transition state matrices are used to segment the trajectory. Processes of development, testing, and reducing the required number of features in transition state matrices are described in detail in this chapter. Data discretization for transition state matrices development can be done in several ways. Three techniques were applied in this research: discretization by the method of equal spacing, equal frequency and k means method. The method for classifying transport modes is described through a random forest algorithm and an algorithm for hyperparameters optimization. Finally, the multilevel method for trajectory segmentation and classification is presented. At the end of the fourth chapter, the test results of the multilevel method on the test problem are presented. Mobile application "Collecty" for collecting data from the mobile device’s sensors, the basic structure and distribution of transport modes in the collected dataset are presented in chapter five. The mobile application contains introductory activities for login and reading the terms of use, the central part for transport mode selection, and the last part for trajectory validation by the user. The collected dataset contains 8 different transport modes and was collected by 15 users. One of the transport modes contained in the dataset is the electric scooter, for which no publicly available data are available in the literature. At the end of the fifth chapter, the validation results of the multilevel method for trajectory segmentation and classification on the collected dataset are presented. In the final chapter, the basic conclusions of the doctoral dissertation are given, and the original scientific contributions achieved through the dissertation are highlighted. At the end of the conclusion chapter, a review of topics for future research is given. |
