Article Preview
TopRecently, many studies have been carried out for trajectory classification. (Silva, Petry, & Bogorny (2019)) summarized three types of trajectory features: global features, local features, as well as global and local features, and presented an experimental comparison for several datasets with methods proposed by others. (Zheng, Li, Chen, Xie, & Ma (2008)) and (Sharma, Vyas, Schieder, & Akasapu (2010)) used global features to classify trajectories. The difference being that (Zheng, Li, Chen, Xie, & Ma (2008)) focuses on the transportation mode classification, so global features extracted by (Zheng, Li, Chen, Xie, & Ma (2008)) and (Sharma, Vyas, Schieder, & Akasapu (2010)) are different. Different studies show different choices for the classifiers. (Mlıch & Chmelar (2008)) and (Bashir, Khokhar, & Schonfeld (2007)) used Hidden Markov Models(HMM) to complete the classification task, while (Wang, Chu, Jiang & Li (2019)) used a Naive Bayesian Model (NBM) and (Liu & Lee (2017)) used BiLSTM. (Xiao, Wang, Fu, & Wu (2017)) used a Multilayer-Perceptron (MLP) and achieved a state-of-the-art accuracy of 75.56% and an F1 score of 73.83% for raw trajectories dataset “Geolife1” and achieved a state-of-the-art accuracy of 93.58% and an F1 score of 93.58% for the raw trajectories dataset “Animals2”, while (Ferrero, Alvares, Zalewski, & Bogorny (2018)) proposed a new method called MOVELETS to achieve the goal of robust trajectory classification. With a Support Vector Machine, the MOVELETS method achieved a state-of-the-art accuracy of 92.30% and an F1 score of 90.82% for the multiple-aspect trajectories dataset “Animals2”. The Gradient Boosting Decision Tree algorithm, which is widely used in many fields, and, in this paper, was proposed by (Friedman, (2001)). (Son, Jung, Park, & Han (2015)) applied the GBDT algorithm to object tracking tasks, and (Rao et al. (2019)) applied it to feature selection problems. At the same time, there are many optimizations for GBDT, such as the LightGBM algorithm proposed by (Ke et al. (2017)). Other optimization algorithms such as (Wang et al. (2018)) and (Wang et al. (2019)) can also be used. Canonical correlation analysis (CCA) was proposed by Hotelling, H., which is fully described in (Hotelling (1935)) and (Hotelling (1992)). There exists an expansion of CCA for a multiset called multiset canonical correlation analysis (MCCA). (Lisanti, Karaman, & Masi(2017)) used MCCA for person reidentification tasks, while (Kanatsoulis, Fu, Sidiropoulos, & Hong (2018)) used MCCA for largescale data, and (Nielsen, (2002)) used MCCA for the GIS system, with discussion of the different restrictions on the MCCA optimization problem.