Autonomous Car: Multi Sensor Data Fusion

Project name: Multi-Sensor Data Fusion for Vehicle Navigation

Prof. Ali Ghaffari

Dr. Alireza Khodayari

Cooperator Researcher(s):

Sina Nosoudi

Sadegh Arefnezhad




Autonomous navigation is one of the most key technologies for driverless cars. Accurate positioning and orientation estimation of vehicles is generally regarded as the basis of many sophisticated modules such as environmental perception, path planning, and autonomous decision-making of driverless cars under complex urban scenarios.
Our main goal is to find a model by data fusion of MEMS sensors and optical encoder. We can find a car position by this sensors data fusion.


Neuro-Fuzzy Adaptive Calibration of MEMS-based Accelerometer for Vehicle Navigation 


Micro-Electro Mechanical System (MEMS)-based inertial sensors have broad applications in moving objects including in vehicles for navigation purpose. The low-cost MEMS sensors are normally subject to high dynamics errors such as linear or nonlinear bias, misalignment errors and random noises. However, the cost reduction of sensors, while keeping their accuracy in a reasonable rang has always been a challenge for engineers. In this paper, a method based on soft computing approaches is presented for calibrating low-cost MEMS accelerometers. Using a fuzzy subtractive clustering algorithm, an initial model for error sources is first identified and then is improved by using adaptive neuro-fuzzy inference system (ANFIS). The calibrated acceleration is then used in a Kalman filter to calculate the vehicle velocity and position. The performance of the presented approach has been validated in real experimental driving scenarios. The results show that this method can improve the accuracy of the accelerometer output, measured velocity and position of the vehicle by 79.11%, 97.63% and 99.28%, respectively. The presented procedure can be used in collision avoidance and emergency brake assist systems.

Calibration of Inertial Accelerometer Using Trained Neural Network by Levenberg-Marquardt Algorithm for Vehicle Navigation


The designing of advanced driver assistance systems and autonomous vehicles needs measurement of dynamical variations of vehicle, such as acceleration, velocity and yaw rate. Designed adaptive controllers to control lateral and longitudinal vehicle dynamics are based on the measured variables. Inertial MEMS-based sensors have some benefits including low price and low consumption that make them suitable choices to use in vehicle navigation problems. However, these sensors have some deterministic and stochastic error sources. These errors could diverge sensor outputs from the real values. Therefore, calibration of the inertial sensors is one of the most important processes that should be done in order to have the exact model of dynamical behaviors of the vehicle. In this paper, a new method, based on artificial neural network, is presented for the calibration of an inertial accelerometer applied in the vehicle navigation. Levenberg-Marquardt algorithm is used to train the designed neural network. This method has been tested in real driving scenarios and results show that the presented method reduces the root mean square error of the measured acceleration up to 96%. The presented method can be used in managing the traffic flow and designing collision avoidance systems.