An Effective Multi-Cue Positioning System for Agricultural Robotics
ABSTRACT
The self-localization capability is a crucial component for Unmanned Ground Vehicles (UGV) in farming applications. Approaches based solely on visual cues or on low-cost GPS are easily prone to fail in such scenarios. In this paper, we present a robust and accurate 3D global pose estimation framework, designed to take full advantage of heterogeneous sensory data. By modeling the pose estimation problem as a pose graph optimization, our approach simultaneously mitigates the cumulative drift introduced by motion estimation systems (wheel odometry, visual odometry, . . . ), and the noise introduced by raw
GPS readings. Along with a suitable motion model, our system also integrates two additional types of constraints: (i) a Digital Elevation Model and (ii) a Markov Random Field assumption. We demonstrate how using these additional cues substantially reduces the error along the altitude axis and, moreover, how this benefit spreads to the other components of the state. We report exhaustive experiments combining several sensor setups, showing accuracy improvements ranging from 37% to 76% with respect to the exclusive use of a GPS sensor. We show that our approach provides accurate results even if the GPS unexpectedly changes positioning mode. The code of our system along with the acquired datasets are released with this paper.
EXISTING SYSTEM :
IT is commonly believed that the exploitation of autonomous robots in agriculture represents one of the applications with the greatest impact on food security, sustainability, reduction of chemical treatments, and minimization of the human effort. In this context, an accurate global pose estimation system is an essential component for an effective farming robot in order to successfully accomplish several tasks: (i) navigation and path planning; (ii) autonomous ground intervention; (iii) acquisition of relevant semantic information. However, self-localization inside an agricultural environment is a complex task: the scene is rather homogeneous, visually repetitive and often poor of distinguishable reference points. For this reason, conventional landmark based localization approaches can easily fail. Currently, most systems rely on high-end Real-Time Kinematic Global Positioning Systems (RTK-GPSs) to localize the UGV on the field with high accuracy. Unfortunately, such sensors are typically expensive and, moreover, they require at least one nearby geo-localized ground station to work properly. On the other hand, consumer-grade GPSs1 usually provide noisy data, thus not guaranteeing enough accuracy and reliability for safe and effective operations. Moreover, a GPS cannot provide the full state estimation of the vehicle, i.e. its attitude, that is an essential information to perform a full 3D reconstruction of the environment.
PROPOSED SYSTEM :
In this paper, we present a robust and accurate 3D global pose estimation system for UGVs (Unmanned Ground Vehicles) designed to address the specific challenges of an agricultural environment. Our system effectively fuses several heterogeneous cues extracted from low-cost, consumer grade sensors, by leveraging the strengths of each sensor and the specific characteristics of the agricultural context. We cast the global localization problem as a pose graph optimization problem the constraints between consecutive nodes are represented by motion estimations provided by the UGV wheel odometry, local point-cloud registration, and a visual odometry (VO) front-end that provides a full 6D ego-motion estimation with a small cumulative drift2. Noisy, but driftfree GPS readings (i.e., the GPS pose solution), along with a pitch and roll estimation extracted by using a MEMS Inertial Measurement Units (IMU), are directly integrated as prior nodes. Driven by the fact that both GPS and visual odometry provide poor estimates along the z-axis, i.e. the axis parallel to the gravity vector, we propose to improve the state estimation by introducing two additional altitude constraints: 1) An altitude prior, provided by a Digital Elevation Model (DEM); 2)
A smoothness constraint for the altitude of adjacent nodes3. Both the newly introduced constraints are justified by the assumption that, in an agricultural field, the altitude varies slowly, i.e. the soil terrain can be approximated by piece-wise smooth surfaces. The smoothness constraints exploit the fact that a farming robot traverses the field by following the crop rows, hence, by using the Markov assumption, the built pose graph can be arranged as a Markov Random Field (MRF). The motion of the UGV is finally constrained using an Ackermann motion model extended to the non-planar motion case. The integration of such constraints not only improves the accuracy of the altitude estimation, but it also positively affects the estimate of the remaining state components, i.e. x and y . The optimization problem is then iteratively solved by exploiting a graph based optimization framework in a sliding-window (SW) fashion , i.e., optimizing the sub-graphs associated to the most recent sensor readings. The SW optimization allows to obtain on-line localization results that approximate the results achievable by an off-line optimization over the whole dataset.
CONCLUSION
In this paper, we present an effective global pose estimation system for agricultural applications that leverages in a reliable and efficient way an ensemble of cues.We take advantage from the specificity of the scenario by introducing new constraints exploited inside a pose graph realization that aims to enhance the strengths of each integrated information. We report a comprehensive set of experiments that support our claims the provided localization accuracy is remarkable, the accuracy improvement well scale with the number of integrated cues, the proposed system is able to work effectively with different types of GPS, even in presence of signal degradations. The open-source implementation of our system along with the acquired datasets are made publicly available with this paper.