Classifying transportation mode and speed from trajectory data via deep multi-scale learning

Abstract

With the rapid development of mobile Internet, the Internet of Things and other new technologies, mobile devices are generating massive amounts of spatio-temporal trajectory data. This paper aims to propose a method that can automatically classify transportation mode and speed, help people understand the mobility of moving objects, thus making people’s life more convenient and traffic management easier. Although there have been some studies on trajectory classification, yet they either require manual feature selection or fail to fully consider the impact of time and space on classification results. None of them can extract features automatically and comprehensively. Hence, we propose Deep Multi-Scale Learning Model and design a deep neural network to learn features under multi-scale time and space granularities automatically. The obtained features are fused to output final classification results. Our method is based on the latest image classification network structure DenseNet, and incorporates attention mechanism and residual learning. This model is able to fully capture spatial features so as to enhance feature propagation and capture long-term dependence. Moreover, the number of network structure parameters is also reduced. We have evaluated our Deep Multi-Scale Learning Model on two real datasets. The results show that our model is superior to the current state-of-the-art models in top-1 accuracy, recall and f1-score. Furthermore, the classification results from our model can help to understand mobility accurately.

Introduction

Nowadays, population explosion has brought many problems to urban traffic management. The growing international trade activities are intensifying port congestion. Identifying and understanding the mobility of moving objects in cities and waters are helpful for managing traffic and meeting peopleâs traveling needs. For example, if we know that taxis are passing quickly on a certain road at a certain time, then we can choose appropriate means of traveling. Moreover, we can infer that the traffic on that road is smooth. Classifying transportation mode is an important method to uncover mobility of moving objects [1] or suspicious vehicle trajectories [2]. Because speed category mining can be seemed as a sub-question of transportation mode classification, this paper will focus on classifying transportation mode and speed. Accurate classification results can ensure that the mobility of moving objects mined from trajectory data is reliable and practical.

Our goal is to classify transportation modes and speed, help people understand the mobility of moving objects, thus making people’s life more convenient and traffic management easier. For instance, given a trajectory dataset in a region and a period, we use trajectory classifier to classify transportation modes and speed of the trajectory dataset. By Summarizing the classification results, we can infer that the objects of the trajectory dataset move at which speed in the region and the period, and can also further understand the traffic conditions.

Although there are some studies on trajectory classification, they mainly focus on the classification of transportation modes, and we want to classify transportation modes and speed at the same time, which can uncover the mobility of moving objects at different time and space granularities. Besides that, they are unsuitable for our tasks because none of them can meet the following two requirements at the same time:

• Extracting features automatically. Traditional machine learning algorithms for trajectory classification normally require feature engineering to extract features, such as motion parameters (speed, direction, accelerated speed) [1], [3], region-based and trajectory-based features [2]. For example, Zhang et. al. classify a trajectory into several traffic modes, compute feature values for each trajectory segment, and finally input these values into decision trees, SVMs, Bayesian networks and conditional random fields [1], [3]. Lee et. al. present a framework for frequent sequential pattern-based classification for trajectories on road networks, while frequent sequential patterns mean frequent combined features composed of single features which are traversed by at least one trajectory [4]. The mentioned above trajectory classification algorithms rely heavily on manual feature selection and cannot work automatically.

• Extracting features comprehensively. Some researchers use deep neural networks to automatically extract features for trajectory classification [5]. Endo et. al. map trajectories into grids, where each grid value indicates the time that the trajectory stays in this grid. Then, they convert the grids into vectors and obtain feature values through DNN. Finally, the feature values and the extracted mobility features are combined into new features which are fed to classifiers for classification [5]. However, their research only classifies transportation modes and fails to describe trajectories from the angles of moving object and speed. Although the mentioned above trajectory classification algorithms based on deep neural networks take into account the impact of space, yet they ignore the impact of time attributes on classification results. Moreover, as the input data to the DNN model is vector data, it is difficult to fully extract local spatial features. Hence, their classification results are less accurate.

To solve these problems, we propose Deep Multi-Scale Learning Model. We first map trajectory data onto girds and then design models for classification. Since trajectory data is mapped onto grids, there is no need to extract features through complex feature engineering. Furthermore, the error of features caused by trajectory point bias can be reduced. Then, we design a deep neural network to automatically learn features so as to fully automate the whole model. Inspired by attention mechanism [6] and the latest network structure [7], [8], [9], [10], [11], [12], [13], we design a Attention Dense Module by integrating attention mechanism, DenseNet [14] and residual learning [15]. Attention mechanism is able to capture long-term dependence and select some major features. The network structure of DenseNet is able to capture local features and spatial features and enhance the propagation of features throughout the network. Consequently, fewer network structure parameters are needed. Residual learning can enable the network to be as deep as possible. Besides, as the original trajectory data is mapped into grid data, different time granularities and space granularities will make difference in grid data, thus influencing classification results. Therefore, we build models for grid data under different time and space granularities, respectively. Finally, we fuse the feature representations outputted by the model as final classification results.

After training and optimizing the model, input the pre-processed real-time trajectory data, and we can get the corresponding transportation mode and speed category. The classification result is summarized, and it is possible to infer at what speed a particular type of moving object is traveling within a certain period in a given area. For example, given an area and a period, we obtain the classification results that a large number of taxis travel at high speed, it can be inferred that the road section is not congested. Therefore, the summary analysis based on the classification results of the model output can answer the questions related to traffic management. for example, (1) How different is the speed of several moving objects in the same location at the same time? (2) In different locations and at different times, what is the speed of the same type of moving objects? (3) With the speed of a given type of moving objects, can we find the time and location where this condition frequently happens? If the output classification results from our model can answer these questions, then it would be helpful for traffic management and personal travels. For instance, our model may be used to predict the time and location of bus congestion. By using this information, we can plan new bus routes and timetables so that the bus transportation system can be more efficient. Our model may be used to determine the speed of taxis in a certain region at a certain time so as help drivers choose a less congested route.

Our contributions are as follows:

• We design Attention Dense Module (ADM) by integrating attention mechanism and DenseNet structure. By stacking ADM modules, we can augment the propagation of features throughout the network and reduce the number of network parameters.

• We propose Deep Multi-Scale Learning Model that can model grid data under different space and time granularities as to fully capture the impact of time and space on classification results.

• We have evaluated our proposed model on two real datasets (Geolife and Ningbo AIS data). Results show that our model surpasses the currently advanced models in top-1 accuracy, recall and f1-score.

The rest of this paper is organized as follows. Section 2 gives definitions and describes our model. Section 3 presents the details of our model. Section 4, several experiments are conducted on real data with our model. Section 5 reviews the related work. Finally Section 6 concludes the whole paper.