Vulnerability of GPS to Provide Vehicle States in Real Time

Vulnerability of GPS to Provide Vehicle States in Real Time

4th IFAC Conference on Modelling and Control in Agriculture, Horticulture and Post Harvest Industry August 27-30, 2013. Espoo, Finland Vulnerability ...

2MB Sizes 4 Downloads 28 Views

4th IFAC Conference on Modelling and Control in Agriculture, Horticulture and Post Harvest Industry August 27-30, 2013. Espoo, Finland

Vulnerability of GPS to Provide Vehicle States in Real Time F. Rovira-Más* *Agricultural Robotics Laboratory, Universidad Politécnica de Valencia, Valencia Spain (Tel: 34-963877291; e-mail: [email protected]). Abstract: The essential technology to take emergent applications, such as precision agriculture and robotics, to actual farming fields is readily available and generally accessible. However, the implementation rate of new technologies has slowed down in recent years and remains low at present. An important cause for this negative outcome is given by the lack of long-term reliability found in global positioning systems. This article analyzes the effects of position errors induced by undetected outliers included in the trajectories of vehicles, exposes the major challenges in the estimation of real-time heading and forward speed, and proposes some practical solutions to strengthen the measurement of vehicle states with global positioning devices. In particular, it is recommendable to move beyond embedded quality indicators by implementing strong filtering routines directly on strategic fields of GGA strings. Dynamic states heading and speed greatly benefit from the efficient combination of VTG-based course and velocity with values directly calculated from a sequence of points determined by GGA geodesic coordinates and time. Finally, whenever a grid-based approach is feasible, the reduction of information offered by a mesh results in more robust field mapping as long as global references are maintained for each cell of the crop map. Keywords: Global positioning systems, Position errors, Position estimation, Yaw estimation, Velocity measurements, Off-road vehicles, Autonomous vehicles, NMEA 0183 strings. implementing a high-quality receiver may suffer from severe signal corruption due to atmospheric errors, multipath reflections, antenna blockage, or simply data transmission jamming (Rovira-Más and Banerjee, 2013). The working conditions typically found in agricultural environments are hard and unpredictable, and therefore it is significantly difficult to grant optimum performance of GPS receivers when operating outdoors and over long working days. To begin with, GPS dynamic accuracy depends on the travel direction due to satellite geometry in mid-latitude areas where satellite distribution is uneven; consequently, the cross-track dilution of precision (XDOP) perpendicular to the travel direction increases as the reference axis changes from 0º to 90º, justifying the direction dependency of cross-track errors (Wu et al., 2006).

1. INTRODUCTION Over twelve years have passed since the US Department of Defense canceled the selective availability for the GPS, encouraging all kind of civil applications that benefit from global referencing. Among them, the nascence of Precision Agriculture (PA) more than two decades ago meant a turning point for the evolution of modern agriculture. Novel applications such as auto-steering, variable rate applications, and crop mapping have revolutionized the way agriculture is understood and managed. However, the rate of adoption for PA-based technology by growers of many crops remains low (Bramley, 2009), and Pölling et al. (2010) have corroborated that the implementation of precision agriculture in practical farming has slowed down in recent years on a global scale. Several factors may account for this negative trend. The complexity of operating computer-based systems by producers hardly ever qualified to handle information technology, in addition to extra costs for unproven benefits, partially justifies these low adoption rates. But the lack of reliability and long-term consistency very likely explains the steady abandonment of PA techniques. Given that global navigation satellite systems (GNSS), GPS in particular, rest on the core of most applications based on agricultural robotics and precision farming, the assurance of stability and accurate measurements from positioning receivers will probably contribute to increase low adoption rates and promote the desired move from concept proofs to commercial products. Unfortunately, the performance of GPS receivers not only depends on the specifications advertised by manufacturers but also on the complete architecture of the localization system. As a result, an onboard system 978-3-902823-44-1/2013 © IFAC

The objective of this article is to identify the principal causes that increase the vulnerability of global positioning systems, and propose actuation protocols to palliate their effects as a way to encourage the practical implementation of precision agriculture by average producers, especially on orchard environments where satellite positioning may result quite challenging. 2. POSITION ERRORS PROVOKED BY OUTLIERS A preliminary measure to enhance the performance of positioning devices consists of enabling receivers to use differential signals (DGPS receivers), as satellite ephemeris and clock errors are cancelled; but unfortunately, multipath and receiver errors cannot be eliminated by differential corrections. Even the sub-inch accuracy reached by real time kinematics systems (RTK-GPS) will be subjected to this type

207

10.3182/20130828-2-SF-3019.00002

IFAC AGRICONTROL 2013 August 27-30, 2013. Espoo, Finland

of errors. Furthermore, it is practically impossible for common users to decouple atmospheric effects, multipath errors, and receiver noise. As a result, the most efficient way to deal with these inaccuracies in field applications is by detecting and eliminating unrealistic data from the primary source of information, i. e., from the data fields that compose GPGGA messages of Standard NMEA 0183 strings, in such a way that wrong information never reaches inner processes. The fields of interest retrieved from GPGGA messages typically are GPS time (s), geodesic position (latitude, longitude, altitude), and quality indices (number of satellites and HDOP). The sequence of points plotted in Fig. 1 represents the trajectory of an agricultural tractor traversing four non-adjacent rows in a vineyard spaced 3 m. The height of the vines was approximately 1.5 m, whereas the GPS receiver antenna was mounted on the cabin roof 3 m above the ground. The path followed by the tractor was recorded with a StarFireTM DGPS receiver (Deere & Co, Moline, Il, USA) processing free signals SF1. This receiver can be set to use more accurate licensed signals SF2, but the fact that most of the producers in the area —if not all– are not willing to pay (≈100 €/month) for precise signals, discouraged the payment of the fee for this research, as the idea was to reproduce field reality as faithfully as possible.

number of satellites for all the points represented in Fig. 1 (Fig. 2a) oscillated between 6 and 10, whereas the HDOP moved from 1.1 to 2.2. Moreover, the detailed analysis of the outliers shows that four of them occurred between points 30 and 60 where the number of satellites was the maximum (10) and the HDOP 1.3. Similarly, other four outliers appeared after point 100, with excellent quality indicators for the satellites. This result verifies that the embedded quality indicators number of satellites and HDOP do not suffice to detect and eliminate dangerous outliers in real time applications. Obviously, those applications that admit off-line processing may filter erroneous estimates any time after the data have been recorded, but when maps and decisions have to be made onboard, ill-determined data need to be removed in real time.

(a)

Fig. 1. Position errors in a tractor’s path along vineyard rows. The trajectory of Fig. 1 clearly shows the presence of 11 outliers, some of them offset by more than 20 m. The time stamp associated to each outlier identified the row it pertained, and therefore its offset error in the north coordinate was easily determined. A customized program coded in C++ acquired NMEA strings from the GPS, retrieved the key fields of information for each point, and recorded all the data in standard text files on a laptop computer (2.2 GHz processor and 1.5 GB RAM) kept inside the cabin of the tractor. The first thought that comes to mind is the coincidence of unfavorable conditions for signal reception, either due to an insufficient number of satellites (below 4) or caused by their uneven relative position in the sky (HDOP greater than 3), or even both situations. Given that every recorded point of the trajectory carries the number of satellites in solution and HDOP, it may seem trivial to discard outliers by directly removing those points from the path by just checking the number of satellites and HDOP. Fig. 2 proves that things are not so straightforward. The

(b) Fig. 2. Number of satellites (a) and HDOP (b) for the trajectory represented in Fig. 1. The evidences found in Fig. 2 indicate that the deterministic assessment of GPS stability is not trivial. Number of satellites and HDOP certainly offer a coarse filter that contributes to the partial removal of outliers but more refined algorithms are necessary to grant full stability. An inner look into the morphology of the basic strings sent by the GPS receiver after long periods of observation revealed severe textual corruption in the principal components of NMEA strings. Fig. 3 provides an example of text degradation at the root of GPS messaging, when the receiver terminal was directly output through the computer monitor in text format rather than connected to the acquisition application. In these cases 208

IFAC AGRICONTROL 2013 August 27-30, 2013. Espoo, Finland

of severe data corruption, strong filters need to be applied upon reception of the original GPGGA strings. Regular GPS receivers work at 5 Hz, therefore it is better to skip uncertain strings than taking the risk of introducing wrong data whose removal might be more difficult later on.

Fig. 3. Degradation of GPGGA strings sent by GPS receivers.

(a)

3. CHALLENGES IN THE GPS-BASED ESTIMATION OF VEHICLE HEADING AND FORWARD SPEED Vehicle heading (course orientation) and forward speed constitute the key dynamic states in the east-north plane. These parameters are essential for navigation, mapping, variable-rate applications, and coordinate transformations between global and local frames. There exist two alternatives to estimate heading and speed from GPS information: the direct reading of vehicle course and speed over ground from GPVTG strings, and the mathematical calculation of both parameters from a sequence of points whose position and time has been determined by GPGGA strings. Specifically, velocity was estimated by averaging traversed length over elapsed time for three intervals within a moving matrix of ten points, and heading was similarly calculated by averaging partial headings (instantaneous course) within a sequence of points that depended on the forward speed: 16 points for speeds over 5 km/h, and 32 points for slow motion under 5 km/h. Fig. 4 illustrates the definition of partial headings for vehicle speeds over 5 km/h. Partial headings were averaged with different weights, as well as the combination of partial speeds within the 10-point matrix.

(b) Fig. 5. Comparison of heading (a) and speed (b) retrieved from VTG messages and calculated from GGA strings. An important issue in the estimation of vehicle states is the strong dependency of GPS heading on the forward speed, especially for agricultural vehicles that travel at low velocities in many tasks. Fig. 6a shows the heading of a tractor moving at different speeds, as recorded in Fig. 6b. Notice the improvement in the stability of the calculations when the vehicle traveled faster than 7 km/h. At these speeds, VTG and calculated states approximately coincide, but careful processing has to be applied for slow motion tasks.

Fig. 4. Heading calculation for speeds over 5 km/h. A priori, the former estimation based on GPVTG strings has the advantage of promptness whereas the latter admits unlimited manipulations to filter noise and enhance the final outputs. Fig. 5 compares the estimation of heading (a) and speed (b) obtained with both methods. Calculated estimates are slightly lagged (less than 2 s. or 10 points) with respect to VTG data, a delay that depends on the size of the sequence used in the computation, usually different for heading and speed. An important advantage of VTG heading is the quick convergence to true values, while calculated course requires several points in the sequence to lock the right orientation.

(a)

209

IFAC AGRICONTROL 2013 August 27-30, 2013. Espoo, Finland

turns where VTG heading has been degraded as a consequence of signal waning. A similar behavior was found for the estimation of velocity (Fig. 7b). Note, however, the positive effect of a tight control on the calculation of heading and speed from a discrete sequence of points, where checking routines implemented along the calculation were capable of measuring the true values of heading and speed. 4. SOLUTIONS AND RECOMMENDATIONS All the information coming from a GPS receiver is transferred through NMEA strings, usually of the type VTG or GGA. VTG messages directly provide course and velocity whereas GGA messages carry time, signal quality indices, and global positioning from which dynamic states may be deduced. Spurious strings such as that shown in Fig. 3 introduce severe interruptions in the data flow and pose important challenges to the stability of the positioning, mapping, or navigation systems. Therefore, the first steps in the enhancement of positioning systems may be directed to parse NMEA strings in order to grant consistency on every significant field. This operation is not trivial because the length and form of some fields depend on the specifications of the receiver used, so that high performance devices tend to feature more digits to increase precision and vice versa. Furthermore, the typical low frequencies at which GPS data are transmitted, either 1 Hz or 5 Hz, sometimes complicate the broadcast of information through the RS-232 port, as many times NMEA strings are delivered incomplete. This source of errors can be improved in practice by acquiring GPS messages in independent but consistent pieces rather than only when GGA strings are complete. Once an entire set of GGA fields is available, further consistency filters can be applied to each field in order to assure that altitudes are realistic for the field used, the number of satellites is always above five to avoid a poor triangulation, the HDOP is less than 3 for a favorable distribution of satellites in the sky, and the forward velocity is physically possible for an agricultural vehicle, normally limited to 40 km/h. As soon as any of these logic conditions is not met, the current point is discarded and a new message is retrieved from the receiver. A conditioning algorithm of these characteristics was successfully implemented in a medium-size tractor to discard outliers and improve global positioning in complex off-road environments (Rovira-Más and Banerjee, 2013).

(b) Fig. 6. Influence of the traveling velocity of the vehicle on the calculation of planar states. According to Standard X587 (ASABE, 2011), a loss of signal is more common in agriculture at the edge of fields, and thus it proposes conducting Dynamic Signal Reacquisition tests on the U-turns. As a matter of fact, VTG measurements have been found to be especially sensitive to headland turns. Fig. 7 shows the heading (a) and speed (b) recorded by an agricultural tractor traversing ten 130-m-long rows of a vineyard.

(a)

The challenges described in Section 3 on the consistent measurement of vehicle dynamic states —heading and forward velocity– have shown that VTG course does not suffer from initial delays or inaccuracies but may be unreliable when the vehicle takes sharp turns to change rows. When dynamic states are directly calculated from position and time carried in GGA messages, the vehicle needs several seconds to find the right orientation, and heading is not available at start-up. On the other hand, the higher controllability of trajectory-based calculations usually results in more stable outcomes over the headlands and other sections of the field where signal reception may be impaired. Overall, the safest solution for real-time applications involves the optimal combination of both sources in such a way that VTG course and speed must have priority at initiation but GGA-based calculated estimates should be dominant in the

(b) Fig. 7. Influence of headland turns on the estimation of VTGbased planar states. In Fig. 7a, VTG-based and calculated heading practically match except for several intervals right after various headland 210

IFAC AGRICONTROL 2013 August 27-30, 2013. Espoo, Finland

vicinity of headland turns or when sharp turns and fast maneuvers are required. The instantaneous availability of both ways of estimating dynamic states offers a favorable degree of redundancy to enhance the deployment of automated off-road equipment. The requirements of precision for GPS receivers onboard farm vehicles greatly depend on the specific tasks assigned to the vehicle. So, for example, precision planting often requires the sub-inch accuracy only reached by RTK-GPS. Yield monitoring or spraying, on the contrary, may accept positioning errors of various decimeters. In addition, even though the 5 Hz sampling rate of usual receivers is not high, the accumulation of data points after several hours of work may be so large that only a grid-based approach can assure a real-time performance. A working session of four hours requiring a continuous acquisition of data to build a map implies the registration of 72000 points with a GPS running at a frequency of 5 Hz. A two-dimensional grid helps to synthesize information in regular cells whose dimension —i. e., the resolution of the map– can be set by the user. Most terrain maps configured as regular grids are based on SLAM techniques (Simultaneous Localization and Mapping), which create or update maps while the vehicle is navigating. This approach is not convenient for agriculture applications where maps must hold consistency in terms of origin and axes over seasons and years. In this situation, therefore, global referencing is indispensable to grant the compatibility of historical data, current data, and even future data. A methodology to combine regular grids and global positioning was developed by Rovira-Más (2012). The algorithm envisioned to do so, has to transform vehicle-fixed coordinates to a global frame defined in the local tangent plane coordinate system. For this operation, the accuracy and reliability of heading is crucial, which in turns requires the right estimation of speed, given the strong dependency of heading calculation on forward velocity. Therefore, although grid maps can lower the accuracy requirements for positioning, the dynamic states heading and velocity still need to be determined with certain precision. Fig. 8 represents a global-referenced navigation grid of several rows of a vineyard, where each cell represents a square of terrain with a side of 25 cm. Filled squares indicate the existence of non-traversable objects (vine canopy), in such a way that the darker the fill color is —therefore larger index on the color bar– the higher chance of detecting an obstacle above ground. The likelihood of a portion of space being occupied by an object (dark cell) was determined by a binocular stereo camera that provided a three-dimensional (3D) scene of the space ahead of the tractor. The fact that canopy rows are rather straight and practically parallel indicates that real-time measured heading was consistent along the mapping mission, which in turns implies that forward speed was correctly estimated too.

Fig. 8. Grid-based approach to palliate navigation errors induced by GPS receivers in real-time applications. 5. CONCLUSIONS The accessibility of GPS information has revolutionized agricultural systems in the last two decades at the same time that the foundation for the agriculture of the future has been settled through robotics and precision agriculture. The actual implementation rates of these technologies, however, have remained excessively low, or inexistent for some crops. This negative outcome may be caused, among other factors, by the lack of long-term reliability of GPS data. This article has analyzed the vulnerability of global positioning systems to provide vehicle states in real time. In particular, severe instabilities have been detected when NMEA messages get corrupted unexpectedly even when the transmission status reports a high number of satellites in the solution and low HDOP values. This fact limits the exclusive use of embedded quality indices to admit or discard GPS data strings, and leads to the implementation of more complex filters to sieve NMEA GGA strings. The estimation of dynamic states heading and speed is favored by the redundant source of information available in usual GPS receivers, either directly from VTG messages, or alternatively from elaborated calculations on GGA position and time. On the whole, global positioning is absolutely necessary for new approaches such as precision agriculture or automatic steering, but GPS information cannot be straightforwardly implemented in practical applications unless effective filtration routines guarantee the transmission of reliable and consistent data. This article has described some of the key issues to check in the implementation of GPS receivers onboard farm vehicles performing advanced tasks, but there may be many other proofs to verify before an intelligent vehicle can execute complex tasks in the middle of a real field for realistic periods of time. Only when farmers, producers, and machine operators have the certainty that modern applications based on new technologies are reliable, consistent, and worth investing their time and money, will these new methods take off to make agricultural production sustainable; and for this to happen, global positioning will have to play a privileged role.

211

IFAC AGRICONTROL 2013 August 27-30, 2013. Espoo, Finland

REFERENCES ASABE (2011). Dynamic testing of satellite-based positioning devices used in agriculture. Standard X587 (Draft 10). Available from http://elibrary.asabe.org. [Accessed on 3 December 2012] Bramley, R. G. V. (2009). Lessons from nearly 20 years of Precision Agriculture research, development, and adoption as a guide to its appropriate application. Crop & Pasture Science, 60(3), 197-217. Pölling, B., Herold, L. and Volgmann, A. (2010). Assessing the potential use of Precision Farming technologies in the EU. In Proc. International Conference on Agricultural Engineering, Clermont-Ferrand, France. Rovira-Más, F. (2012). Global-referenced navigation grids for off-road vehicles and environments. Robotics and Autonomous Systems, 60, 278-287. Rovira-Más, F. and Banerjee, R. (2013). GPS data conditioning for enhancing reliability of automated offroad vehicles. Journal of Automobile Engineering, 227 (4), 78-92. Wu, C., Ayers, P. D. and Anderson, A. B. (2006). Influence of travel direction on GPS accuracy for vehicle tracking. Transactions of ASABE, 49(3), 623-634.

212