Multi-models machine learning methods for traffic flow estimation from Floating Car Data

Highlights

• Application of Machine Learning based model to rebuild the traffic flow relationship.

• Construction and training of multi-models to reduce estimated traffic flow error.

• Demonstration of Gaussian Process Regressor to achieve the lowest error.

Abstract

Traffic flow measurement is very important for traffic management systems. However, the existing traditional measurement approaches are highly time-consuming and expensive to continuously gather the required data and to maintain the corresponding equipment, such as loop detectors and video cameras. On the other hand, many services on the web propose to estimate automobile travel time taking into account traffic conditions thanks to crowd sourced data (Floating Car Data). This work proposes to reconstruct, from estimated travel time, traffic flows using machine learning method. In particular, we evaluate the capacity of Gaussian Process Regressor (GPR) to address this issue. After obtaining estimated travel time on a given route, a clustering process shows that travel duration profiles in each day can be associated to different “types of day”. Then, different regressors are trained in order to estimate traffic flows from travel duration. In the “multi-model” variant, we trained a Regressor for each type of day. Conversely, in the “single model” variant, only one Regressor is trained (the type of day is not taken into account). This is an innovative work to estimate and reconstruct the traffic flow in transportation networks with machine learning method from aggregated Floating Car Data (FCD). A series of experiments are conducted to compare the estimated traffic flows, obtained by the proposed single model and multi-model, and the real ones from actual sensors. The obtained results show that both single model and multi-models can capture the tendency of real traffic flows. Furthermore, the performance can be improved by regulating parameters in GPR machine learning model, such as half width of sample window and sample size (a whole week or only weekdays), and multi-models can highly increase the performance compared with the single model. Therefore, the proposed GPR machine learning and FCD based new method can replace those traditional loop detectors for the measurement of traffic flow.

Introduction

With the rapid growth of urban centers during the last decades, the development of efficient urban transportation services has become a central issue to reduce the high wasted time during the daily commute. The resulting increasing demand in terms of transportation flows has to cope with the difficulty to adapt existing or create new transportation networks.

In this context, simulate daily transportation behaviors allows operators to experiment and to visualize decisions about infrastructure and regulation policies. One of the major basics on efficient simulation relies on the ability to produce models representing the way that transportation flows evolve with time, depending on the traffic demands and events that impact the transportation network. The estimation of traffic flow is one of the core requirements in those simulation. One of the costless solution would be to reconstruct the traffic flow from aggregated information (travel duration estimations), available on web services.

Previous work validates that machine learning based approach is a promising way to reconstruct ”sensor like traffic flow data” from aggregated information like the ones proposed by Google services (Li et al., 2019a, Li et al., 2019b). Applied such an approach would permit, for instance, to infer on a realistic traffic demand at each entrance of a city (i.e. the flow of incoming vehicles).

Machine learning over aggregated information approach is based on accessible databases that can provide information regarding the transportation condition (travel duration) at a given location and at a given time. In 2007, the Google company has extended Google Maps by adding Google Live Traffic, the visualization of traffic information in real time (van den Haak et al., 2018, Jeske, 2013). Here, the notion of real time means the current state and is applied to qualify the service of FCD provided. In more detail, Google exploits users’ position data of Android smart-phones, in order to get a significantly fast and accurate mapping of the traffic. This data is called Floating Car Data (FCD), which can also be collected by any localization system embedded in a car and sent to the service provider via a mobile connection. Generally, these raw Floating Car Data are aggregated to provide more intelligible and relevant information regarding traffic condition. For example, in Google Maps, FCD are used to give a real-time traffic information using colored road section¹ which is determined by the navigation system or, as in the case of Google Live Traffic, by the smart phone and is sent to the service provider via a mobile phone connection. Therefore this allows the generation of real-time traffic information, which is visualized by the colors on Google Maps: red road points are related to a traffic jam or stop-and-go traffic, orange indicates heavy traffic and green points correspond to clear roads. However, those platforms generally provide only aggregated data like average travel duration more than the initial raw data.

Such an approach would permit operators to provide efficient global information while limiting the effort in continuously measuring road traffic flows based on physical sensors (radars, induction loops, etc.). The number of sensors, even in mid-sized cities, can increase very quickly. For example, to measure the input and output flows of a simple 4-points roundabout, at least 8 physical sensors are required. Therefore, such an expensive traffic flow measurement method makes the mentioned simulation process out of reach.

Preliminary result was published on the 15th World Conference on Transport Research 2019 (Li et al., 2019a) and on the 6th International Conference on Control Decision and Information Technologies 2019 (Li et al., 2019b) where traffic flows is estimated according to Floating Car Data (FCD) from Google Maps only on the basis of regressors trained using machine learning techniques instead of using stationary physical equipment (such as loop detectors (Cheung et al., 2005) or video cameras (Coifman et al., 1998)). In this paper, we present an extension of the previous works with an increased experiment setup that permits to automatize the model definition. Firstly, we show experimentally that, among 19 types of regression methods (including Linear Regression Models, Regression Trees, Support Vector Machines and Ensemble of Trees), the Gaussian Process Regression (GPR) is the most suitable machine learning method to obtain the best fitting criterion with respect to our dataset. Secondly, a selection of the adequate regressor is computed from a set of regressors (multi-model) to estimate traffic flows from FCD, based on the different types of travel duration profiles. This multi-model approach can greatly reduce the estimation error, by precisely clustering days presenting different types of travel duration profiles. Experiments are conducted by comparing estimated flows with real ones provided from induction loop sensors. The results we obtain seem promising enough to say that correct transportation flows models could be obtained with a very light use of real traffic sensors.

This paper is organized as follows: the next section describes related works and the different usages of aggregated FCD in the context of transportation networks modeling. The third section focuses on the problem we choose to address that is building sensor like data flow measurements from aggregated data. In this section, we also provide details regarding our problem formulation on the standpoint we took for solving it. The fourth section presents the single and multi-GPR machine learning method for the estimation of traffic volume. The fifth section deals with the experimental site, the results we obtained, the comparison between estimated traffic flow and real observed data prior to the discussion. The last section concludes this paper and presents some perspectives and further works based on it.