Dual Supervised Autoencoder Based Trajectory Classification Using Enhanced Spatio-Temporal Information

Abstract

With the rapid development of mobile internet and location awareness techniques, massive spatio-temporal data is collected every day. Trajectory classification is critically important to many realworld applications such as human mobility understanding, urban planning, and intelligent transportation systems. A growing number of studies took advantage of the deep learning method to learn the high-level features of trajectory data for accurate estimation. However, some of these studies didn't interpret spatiotemporal information well, more importantly, they didn't fully utilize the high-level features extracted by neural networks. To overcome these drawbacks, this paper utilizes the proposed stop state and turn state to enhance spatial information, and at the same time, extracts stronger time information via Recurrence Plot (RP). Moreover, a novel Dual Convolutional neural networks based Supervised Autoencoder (Dual-CSA) is proposed by making the network aware of Predefined Class Centroids (PCC). Experiments conducted on two real-world datasets demonstrate that Dual-CSA can learn the high-level features well. The highest accuracy of the Geolife and SHL datasets are 89.475% and 89.602%, respectively, proving the superiority of our method.

POPAyI: Muscling Ordinal Patterns for Low-Complex and Usability-Aware Transportation Mode Detection

“…To avoid using these hand-crafted features, some studies employ supervised [2,5,6,[20][21][22][23] and semi-supervised [24][25][26] DL techniques to extract multiple layers of features automatically, often yielding comparable or even superior results than ML. Still, they demand significant computational resources and large volumes of training segments with equal length, which require interpolation or padding in real-world data.…”

“…For example, consider the points in Figure 3 within the first sliding window s 1 = {(3, 2), (1,6), (3,6)} and using an amplitude threshold of q = 5. In this case, the Euclidean distance is 4 and the resulting pattern is π 1 = (021, 0).…”

“…Most TMD literature relies on hand-crafted features (reaching hundreds of features [1,3,4]) or computationally-and data-intensive DL methods [5,6], which pose challenges for resource-constrained Internet of Things (IoT) scenarios [2,7]. We tackle TMD differently by leveraging mobility analysis with Ordinal Patterns (OP).…”

“…In addition, compared to only the best CNN in the ensemble, POPAyI leverages F1 score and accuracy by about 10%. Furthermore, increasing the trace size enhances the performance of POPAyI allowing it to achieve similar metric results to Lu and Xia [6] and James [21] without increasing the number of parameters, which is impossible with DL methods. The ability of POPAyI to utilize longer segments provides richer features, enhancing its effectiveness in TMD.…”

Addressing robust travel mode identification with individual trip‐chain trajectory noise reduction

“…In the study of travel mode identification, the importance of TNR has not been paid attention to. For example, when using the Geolife dataset collected by Microsoft Research Asia in Beijing [9,15], some studies [15][16][17][18][19] identify modes on the original trajectory. Zhu et al [20] used linear interpolation to complete positions between records with too large time intervals.…”