An application of a fuzzy classifier extracted from data for collision avoidance support in road vehicles

Abstract

Road traffic collisions are an outstanding problem in current developed societies. This paper presents a solution to support collision avoidance based on the timely detection of the vehicle maneuvers. Since the longitudinal interaction among vehicles, with the commonly known car-following behavior, is one of the most important causes of crashes, it was decided to focus on longitudinal maneuvers, identifying the maneuvering states of cruise, accelerating or decelerating and stop. The classification is carried out by means of fuzzy rules extracted from navigational data. Therefore, in our proposal no extra sensors are needed apart from two commonly installed for navigation purposes: the odometry of the vehicle and an accelerometer. The system was tested with low-cost sensors showing good results when compared to the literature of the field.

Introduction

The detection of hazardous situations in the road has been a focus of large attention from the research community in the last years. Among the possible approaches to the problem, some authors dedicate their work to the so-called cooperative systems, in which the vehicles avoid crashes by means of the exchange of relevant information within their environment (other vehicles and the elements of the infrastructure). For the readers interested in the topic, a relevant project on cooperative systems is the CVIS project (ERTICO, 2009) for Cooperative Vehicle Infrastructure Systems.

The work presented in this paper focuses on the detection of the vehicle maneuvers starting from measures collected by the onboard sensors. Cooperative systems may benefit from this work in such a way that timely detections of the vehicle maneuver may increase the situation awareness of the road scene. As it is proved in the literature, the timely predictions of the vehicle maneuvers are useful for scene interpretation and support collision avoidance (Terroso-Saenz and Valdes-Vela, 2010, Toledo-Moreo et al., 2010, Toledo-Moreo et al., 2006, Toledo-Moreo and Zamora-Izquierdo, 2009b, Zamora-Izquierdo et al., 2007, Zong et al., 2009). Additionally, it could allow improvements in the quality of the position of each vehicle given that the proper identification of the vehicle maneuver could raise the accuracy and robustness of the position estimates. This is due to the fact that, when the position is being estimated, the model that represents the vehicle motion is usually simplified and does not represent equally well every maneuvering mode (Ndjeng et al., 2008). For instance, the movements carried out by a vehicle when it is parking differ significantly from the ones done while driving a highway at constant velocity. To cope with this estimation, the literature shows that it is more efficient to use a finite set of models dedicated to represent the different maneuvering modes (constant velocity, constant acceleration/deceleration, and stationary model), rather than a single highly complex model that must consider every possible situation (Blackman and Popoli, 1999). Therefore, if we can identify the maneuvering mode at a given time t, then we can select the model that matches the best the nature of the vehicle motion at that instant t. A number of publications in this field show that it is possible to improve the position accuracy and its robustness – fail tolerance – following this approach with a set of very few motion models (Ndjeng et al., 2007, Schubert and GerdWanielik, 2009, Toledo-Moreo et al., 2007). The potential improvements in the estimation of the vehicle position are, therefore, a consequence of the maneuver detection and not a direct benefit of our method. For this reason, this aspect is not covered in this paper.

Since the detection of the maneuver of the vehicle is based on measurements of sensors commonly installed onboard the vehicle, no extra costs are added to the system, apart from a certain – reduced – demand of computation. Given that we focus on longitudinal maneuvers and the vehicles are assumed to move on its local tangent plane, only the odometry and an accelerometer aligned with the longitudinal axis are needed. Although positioning systems based on odometry and inertial sensors may suffer from estimate drifts due to the accumulative errors in absence of absolute positioning updates, it is important to notice that input variables of the proposed model are velocity (from the odometry) and acceleration (from the accelerometer), and therefore no accumulated error is introduced. For localization purposes, a GPS receiver installed onboard the test vehicle provides absolute values of position, minimizing in any case the possible consequences of dead-reckoning errors in the vehicle position estimates.

The noisy and imprecise nature of the data commonly returned by low cost vehicle sensors, led us to consider the use of fuzzy rules (Zadeh, 1975). Fuzzy rules have good skills to deal with this kind of data. Furthermore, given that the set of longitudinal maneuvers a vehicle can perform is preestablished (cruise, acceleration or deceleration and stop), the problem of selecting the most suitable motion model at the current instant can be seen as a classification task. Therefore, given an input tuple composed of data coming from the previously mentioned sensors, the goal of the set of fuzzy rules is to classify this tuple into one of the existing maneuvers (classes). Once the maneuver is selected, the most appropriate motion model can be easily applied. This latter step can be done following (Toledo-Moreo et al., 2010), and it is out of the scope of this paper.

The proposed fuzzy classifier has the next features:

• Instead of using expert knowledge, the proposed one is extracted from data. Hence, it has being obtained following a fuzzy modeling approach (Babuka, 1998).

• The input variables of the fuzzy classifier are the current instant longitudinal velocity and acceleration. Then, the proposed fuzzy classifier is independent from both past sensor measures and past inferred maneuvers. Consequently, it does not need to keep buffered information, which leads to a subsequent saving of memory.

• The fuzzy modeling process has been carried out with the only goal of obtaining an accurate fuzzy classifier with no intention of being interpretable for humans. As it will be seen, in the light of the comparison with other proposals in the literature, this accuracy goal has been achieved. Besides, the fuzzy classifier is compact, having a low number of input variables and rules. As a result, little computation time is consumed to infer the output.

• Finally, it must be remarked that a simple classifier is advisable given that it has been focused on longitudinal maneuvers. Therefore, an increase of the model complexity is foreseen when the task of transversal maneuver will be undertaken. In that sense, the simpler the longitudinal maneuver model, the simpler the transversal+longitudinal one. Besides, simple models are better in order to incorporate expert knowledge and to manually adjust their parameters.

The rest of the paper is organized as follows: in Section 2 the most relevant related works are shown. In Section 3 the fuzzy classifier is introduced. Afterwards, in Section 4, the fuzzy modeling process is described. Section 5 shows and analyzes the results of the fuzzy modeling experiments. Finally, the main conclusions are presented in Section 6.