Previous Abstract Return to Session A6 Next Abstract

Session A6: Adaptive KF Techniques, Data Integrity, and Error Modeling

Driver Behavior Assessment based on Loosely Coupled GPS/INS Integration in Harsh Environment
Oussama Derbel, Mohamed Lajmi Cherif and René Jr. Landry, LASSENA/ETS, Canada
Location: Big Sur

Commercial location-based services use mainly the Global Positioning System (GPS) receiver data to develop the driver assistance and monitoring systems. Nevertheless, the GPS has a lot of problems, and the most important one is the signal quality in harsh environment such as urban canyon, tunnel and bridge. To overcome this drawback, the Inertial Navigation System (INS) comes to aid the GPS receiver using Kalman filtering to compute an accurate position, velocity and acceleration in these environments. The basic Kalman filter deals with linear problems while the Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF) and Particle Kalman Filter (PKF) handles non-linear problems. The EKF uses the linearized measurement, the state equations and the Jacobian matrices. So, the choice between these types of Kalman filter depends on the problem constraints and raises the question on how to choose the best filter for insurance applications. Loosely-coupled integration allows the independence and redundancy of GPS and INS solutions to be maintained in addition to provide a more robust coupled navigation solution. Indeed, integration by loosely coupling has a closed loop architecture which allows the correction of certain error parameters of the INS system. The presented architecture uses a Kalman filter to couple the GPS and INS solutions (position and velocity) in order to estimate the position, velocity and attitude errors of the INS solution. It illustrates also the typical architecture of INS with its related components. The INSs are fully autonomous systems and able to compute the position, velocity and attitude of the vehicle based solely on the linear acceleration and angular rates given by the inertial sensors (gyroscopes and accelerometers). The data from the Inertial Measurement Unit (IMU) provides the specific forces, in the body frame, which are transformed to the North-East-Down (NED) frame. The components of force of gravitational attraction, of centripetal force and of acceleration of Coriolis are removed from the vector of specific forces in order to obtain only the acceleration of the mobile according to the navigation frame. The Gravitational and Coriolis accelerations can be estimated using mathematical models as function of the vehicle position and velocity. Then, the acceleration is integrated one time to obtain the velocity and a second time in order to have the position of the mobile. Thus, this model is used to estimate nine error states of the navigation system which are the position error, the attitude error and the velocity error. The error propagation model for GPS/INS integration by loosely coupling is based on the Psi angle error propagation model.
This paper matches the two complementary areas which are: GPS/INS integration and Driver Behavior Assessment (DBA). In the literature, these two fields have been deeply investigated separately. However, an accurate analysis of the driver behavior requires precise data (position, velocity and acceleration, etc.). This paper presents a new method for driving behavior assessment based on the loosely coupled GPS/INS integration that allows a precise results, especially in case of GPS outages which can also be modeled in the driver behavior assessment part.
Traditional methods of DBA use the Hidden Markov Model (HMM) and Neural Network (NN) to estimate the driver behavior. This paper presents a new approach to manage the different aspect of the driving behavior estimation by adopting two levels of information fusion matched by the use of the Fuzzy Theory (FT). The first level is designed to compute locally the risk of related to the Driver, Vehicle and Environment (DVE). The second level fuses theses local risk levels to compute a global risk of driving. The risk related to each parameters of each entity of the DVE system is modeled using FT. This technique is an effective way to improve the computation time and reduce the complexity of the fusion problem. The information fusion is based on the Belief Theory (BT) which is called also Dempster-Shafer Theory (DST) of evidence. The advantage of using this theory is its capability to characterize the uncertainty of the results through the upper (Plausibility) and lower (Belief) bound of the probability of an element in a subset. This probability is also called the Basic Probability Assignment (BPA), which is determined using fuzzy measurement in this paper and designed to model the risk related to each parameter of the DVE system as well as the GPS outages. In fact, the relationship between fuzzy and belief measurement plays a crucial role to reduce the computation time and offers the possibility to model the risk involved by each parameter.
In this paper, the risk of an event is qualified over the sets of the referential subset, given by {Empty, LR , LR U MR, MR, MR U HR,HR,HR U LR,Theta} where LR, MR, HR designate the ”Low Risk”, ”Medium Risk” and ”High Risk” respectively, Empty represents the conflict between sources and Theta the ignorance of the risk (the union of all hypotheses of risk). The determination of the BPA is the second step of the Belief theory. This task is the hardest one in this theory. In this paper, the Fuzzy theory is used to compute the BPAs for two reasons: the first one is the availability of the statistics that can be used to develop the fuzzy model. The second reason is that the fuzzy theory can be used to reduce the number of inputs in each entity of the system Driver Vehicle Environment (DVE) and then it reduces the complexity of the fusion problem when using the Belief theory.
The risk related to the Environment entity is based on the position of the vehicle and the time, day and month of driving. These parameters are given from the GPS and updated in every time step. The position of the vehicle (longitude and latitude) is used to determine the district of driving based on the Open Street Map data. An algorithm has been developed to determine the district, the street and the maximum allowed velocity from the GPS data at each time step for each given position. The drawback of using OSM is the insufficient number of the available data for some cities (e.g. Montreal). Then, the driving street and the maximum allowed velocity are unavailable. To overcome this problem, an interpolation function has been developed. It adds a number of points (positions) to the OSM database between each pairs that are considered as references points and defines the limits of each street. To optimize the processing time, only the box that contains all the positions data we have during the test and given by the GPS is used.
The experimental test contains two parts in this paper. The first one is used to validate the proposed INS/GPS loosely coupled algorithm by inserting GPS outages in a sequence of data. The performances analysis has shown the capability of the proposed algorithm to reduce the errors in terms of position and velocity in case of GPS outages. In fact, the developed loosely coupled GPS/INS integration algorithm determines the vehicle position in tunnel using the INS only. Once the GPS receiver signal is available and the GPS fix value is greater than 3, the Kalman filter updates the errors and determines the new position and velocity of the vehicle.
The error analysis cannot be complete without focusing on the cumulative error in terms of position during the mission. In a first scenario, we have introduced two periods of GPS outages with two different velocities (low speed and high speed). The outage period #1 is in highway where the outage period #2 is characterized by low velocities and long stopping times, and lasts longer than the outage #1. Although the GPS generates a large position error when the vehicle is in the stop mode, the developed loosely coupled algorithm is able to compute an accurate position.
The second part is used to test the developed risk models and analyse the driver behavior as well as the loosely coupled INS/GPS integration algorithm in case of two scenarios. The first scenario considers a sample data and then the evaluation of the driver behavior is made microscopically. We assume that the 32 years-old male drives at 4pm on Monday in May 2016 in the ’Ville-Marie’ district of Montreal (Canada). According to the developed driver risk model, based on his age and gender, the Driver entity distributes the masses over the propositions MR (m(MR)=0.3) and LR U MR (m(LR U MR)=0.7). The ’Ville-Marie’ district is qualified by a medium risk based on the statistics of the number of accident. So, the masses related to the driving place assigns the total mass to the medium risk proposition MR (m(MR)=1). Nevertheless, the masses related to the hour of driving and the month of driving are given to the high risk proposition HR (m(HR)=1) while the masses related to the day of driving is totally assigned to the MR U HR proposition (m(MRUHR)=1). The GPS signal is unavailable since the vehicle is in the tunnel. In this case, the mass related to the Vehicle entity is totally assigned to the proposition Theta. In fact, we consider that there is a risk but we cannot determine it since the velocity and position of the vehicle are unavailable. So, here comes the advantage of using the DST. Nevertheless, using the loosely coupled INS/GPS integration algorithm, we are able to assess accurately the driver behavior in tunnels.
The second scenario is performed by a 32 years-old male driver who drives at 3pm in Montreal (city in Canada). The trajectory contains two GPS outages. The experimental test vehicle Dodge 2012 is equipped with a SPAN Novatel which contains a high precision INS and GPS receiver to collect data (e.g. position, velocity, acceleration, etc.). We used these high precision sensors to validate and study the impact of the GPS outages on the evaluation of the driver behavior intended to serve the insurance companies. Our developed risk models have showed a high sensitivity against the risk parameters at the local fusion level as well as the global fusion levels by studying the variation of the obtained masses in cases of a single sample and a group of data (i.e. mission).
This paper has clearly shown the impact of GPS outages on the evaluation of the driver behavior. It has presented a GPS/INS integration system followed by a Driver Behavior Assessment (DBA) system. Summing up the results, it can be concluded that the DBA has been sensitive to the GPS outages. In addition, the most important finding of this paper is the validity of the developed risk models of the DVE system. This research work can be used in case of insurance application



Previous Abstract Return to Session A6 Next Abstract