Fusing Optimal Odometry Calibration and Partial Visual Odometry via A Particle Filter for Autonomous Vehicles Navigation
Abstract
Autonomous vehicles are increasingly becoming ubiquitous in the 21st century; they find application in agriculture, industry, airplanes, cars, service robotics, and others; in order to display autonomous guidance, a vehicle needs to estimate its position and orientation relative to an arbitrary coordinate system; to do so, several sources of information can be used, including images, global positioning systems, inertial measurements or odometry, each according to the application; methods, such as Kalman Filter can be used to combine the several sources of information; however, the more accurate each source of information is, the better the estimation of vehicle position and orientation will be; therefore, the calibration of the parameters of the odometrical systems in autonomous terrestrial vehicles is a must; visual guidance is also an important technology used for vehicle guidance. In this paper, it is presented an off-line method for odometry calibration using a genetic algorithm and the fusion of odometry data with heading information from camera data; a particle filter is used to fuse the data from the optical encoder and the camera. This method was tested in an Automated Guided Vehicle (AGV) with tricycle topology, demonstrating high accuracy in position estimation and guidance through arbitrary paths.
Keywords
Download Options
Introduction
Autonomous vehicles are a relatively new technology with hundreds of potential applications in many aspects of the human life [1]; they have the potential of becoming the everyday driver of the people, and also the automatic guided vehicles AGV of the new generation industrial plant, not requiring magnetic tracks under the floor to follow a predetermined path; they exhibit the capability of interpreting data from sensors to determine their current position with respect to a predefined coordinate frame, responding at any time to the question where am I? [2]; the pose of a vehicle is comprised of the x,y coordinates of its position plus the heading or yaw angle [x, y, Ɵ]. When two different sensors provide information from the same variable, it is necessary to decide at which extent one is more reliable than the other in order to provide a weighted estimate of the variable; such is the working of the Kalman Filter [3]. Sources of the same variable are, for instance, an inertial measurement unit (IMU) [4] providing yaw rate information -which integrated over time provides yaw angle- and optical encoder ticks which can be counted and the count converted into the yaw angle of the vehicle whose wheels they are attached to.
A navigation strategy, or algorithm, is usually required to move the vehicle from one point to another; the navigation algorithm is fed with the current position and a next desired position, and its outputs convey the commands to motors in charge moving and steering the vehicle. Again, the accuracy of the movement of the vehicle depends on the accuracy of the pose estimation; this leads us to the conclusion that among the first work to be done is to accurately estimate the pose of vehicle in order to provide reliable autonomous guidance.
Conclusion
This paper presents an approach to the problem of autonomous vehicle guidance; the use of optimization in a first stage of odometry calibration is one of major steps towards a real-world applicable system because it helped to reduce systematic and non-systematic errors, as seen in Table 2.
Combining data from the odometry system and the camera with the particle filter, as shown in Figure 2, significantly supported the dramatic reduction in the percentage of error shown as compared to the simple use of odometry calibration. Although many sophisticated algorithms have been published in the past and recent years, including auto-calibration and pose, they are complex and computational expensive, a key difference with the approach presented here. It is not discarded as future work to include some continuous optimal calibration method during the operational stage of the vehicle.
Finally, the goal of this project is to produce industrial grade AGV autonomous systems, so the efforts in such direction will continue from these research groups.
Combining data from the odometry system and the camera with the particle filter, as shown in Figure 2, significantly supported the dramatic reduction in the percentage of error shown as compared to the simple use of odometry calibration. Although many sophisticated algorithms have been published in the past and recent years, including auto-calibration and pose, they are complex and computational expensive, a key difference with the approach presented here. It is not discarded as future work to include some continuous optimal calibration method during the operational stage of the vehicle.
Finally, the goal of this project is to produce industrial grade AGV autonomous systems, so the efforts in such direction will continue from these research groups.