Sensor fusion algorithms for autonomous vehicles

Abstract

Autonomous vehicles are an emerging area with many problems without a definitive solution. One of these problems is the vehicle’s pose estimation which until these days still an open problem. In this research we will propose a sensor fusion strategy that estimates a vehicle’s position and orientation, this algorithm is used for the higher-level tasks of trajectory generation and path following. The VO algorithms that were tested during the research are ORB SLAM2, and the proper VO of the ZED camera by Stereolabs ®. The sensor fusion algorithms that were tested during the research are: Kalman filters, averaging measurements and Neural Networks. The main odometry algorithm was designed in order to function with a GPS, VO, IMU, steering sensor and a speed sensor. This algorithm is compatible with ROS in order to exchange data with sensor and other algorithms with nodes on the ROS network. This and the sub-algorithms that are executed in order to obtain a reliable and accurate pose estimation were tested separately in simulation and experiments and results are included in the results section. The proposed odometry algorithm is used by higher-level algorithms in order to make a simulated Unmanned Ground Vehicle locate itself and navigate to a desired final point, following a previously generated trajectory. Once algorithms were debugged and tested, their behaviour will be observed in an experimental environment using the real UGV designed by our research group in the Tecnologico de Monterrey Campus Puebla. Physical restrictions on the real UGV made the author modify the simulated algorithms in order to work in the real world. These modifications mainly appear in the navigation control law and its behaviour can be found in the Experimental results section. Furthermore links to videos where the behaviour of the UGV, its generated trajectory and some statistics, are included in the experimental results sections.