“We kept the design of the enhanced SPAN-CPT identical to our legacy product to ensure a seamless upgrade process for our customers who would like to take advantage of the improved positioning capabilities,” said Jason Hamilton, NovAtel director of marketing. “The enhanced SPAN-CPT is fully backwards compatible with the previous generation of product. It retains the same compact form factor with identical pin-out and log structure.”

As with the previous generation product, the upgraded SPAN-CPT integrates NovAtel’s precision receiver technology with fiber optic gyro and MEMS accelerometer inertial components from KVH Industries in one compact unit. The tight-coupling of the GNSS and INS technologies optimizes the raw GNSS and IMU data, delivering a superior position, velocity and attitude solution, NovAtel said. Comprised entirely of commercial components, the SPAN-CPT minimizes the operational complexities of working across international boundaries.

Production of the OEM6 supported SPAN-CPT begins June 1.

Spatial Dual contains the Trimble BD982 GNSS receiver, which is a triple frequency dual-antenna RTK GNSS receiver. Using dual-frequency moving baseline RTK, Spatial Dual is able to provide heading accuracy of less than 0.1 degrees using its dual antennas. The dual-antenna heading works while both stationary and moving and allows for very accurate heading in both slow moving and 3D vehicles, where equivalent single antenna systems must rely on magnetic heading. An additional benefit of the dual antennas is the ability to measure slip angle to within 0.2 degrees.

Spatial Dual supports all of the current and future satellite systems, including GPS, GLONASS, Galileo and BeiDou. In addition, Spatial Dual supports RTK for centimeter positional accuracy and the recent Omnistar G2 network for 10 centimeter accuracy.

Spatial Dual provides position, velocity and orientation at rates up to 1000 Hz for highly dynamic applications. When Spatial Dual loses a GNSS fix it continues to navigate using dead reckoning inertial navigation to provide seamless navigation data through tunnels and other outage situations.

Spatial Dual is housed in a precision marine-grade aluminum enclosure that is waterproof and dirtproof to the IP67 standard and shockproof to 2000g, allowing it to be used in tough conditions.

Spatial Dual supports a wide range of peripherals including odometers and wheel speed sensors for ground vehicle navigation, DVLs and USBLs for underwater navigation and many other external sensors. It supports both industry standard NMEA output and an efficient binary protocol.

Applanix AP50 GNSS-inertial system.

The Applanix AP50 GNSS-inertial system is a GNSS-inertial sensor plus inertial measurement unit (IMU) in a compact form factor. It features a high-performance precision GNSS receiver and the Applanix IN-Fusion GNSS-inertial integration technology running on a powerful, dedicated inertial engine (IE) board.

Riegl’s VQ-820-G airborne laser scanner.

On board an unmanned aerial vehicle (UAV), the system is capable of penetrating areas that may be too dangerous for piloted aircraft or ground patrols. This can provide additional safety and security for its users.

“We really appreciate the professional and amicable cooperation with Applanix, which allows us to offer user-friendly and powerful, fully integrated solutions for dynamic data acquisition to the marketplace,” said Jürgen Nussbaum, Riegl director of international sales.

In addition, Applanix will be a Gold sponsor at Riegl LIDAR 2013, Riegl’s international user conference taking place in Vienna, Austria, June 25-27.

The prototype can guarantee the position of the vehicle to within 1 or 2 meters in urban settings, the researchers said.

The system can be installed in any vehicle for little cost and may eventually work on smartphones, the researchers said. Their findings are described in the report, “Context-Aided Sensor Fusion for Enhanced Urban Navigation.”

Sensor Fusion. The prototype system incorporates a conventional GPS signal with those of other sensors (accelerometers and gyroscopes) to reduce the margin of error in establishing a location. “We have managed to improve the determination of a vehicle’s position in critical cases by between 50 and 90 percent, depending on the degree of the signals’ degradation and the time that is affecting the degradation on the GPS receiver,” said David Martín, a researcher at the Systems Intelligence Laboratory (LSI – Laboratorio de Sistemas Inteligentes) at UC3M. The system was jointly designed and developed by LSI and the Applied Artificial Intelligence Group (GIAA – Grupo de Inteligencia Aplicada Artificial).

The margin of error of a commercial GPS, such as those that are used in cars, is about 15 meters in an open field, where the receiver has wide visibility from the satellites. However, in an urban setting, the determination of a vehicle’s position can be off by more than 50 meters, due to the signals bouncing off of obstacles like buildings, trees, or narrow streets. In certain cases, such as in tunnels, communication is lost, hindering the GPS applications reaching Intelligent Transport Systems, which require a high level of security.

“Future applications that will benefit from the technology that we are currently working on will include cooperative driving, automatic maneuvers for the safety of pedestrians, autonomous vehicles or cooperative collision warning systems,” the scientists comment.

Integration of GNSS antenna of rover receiver and IMU in a platform over the roof of the vehicle.

The greatest problem presented by a commercial GPS in an urban setting is the loss of all satellite signals. “This occurs continually, but commercial receivers partially solve the problem by making use of the urban maps that attempt to position the vehicle in an approximate point,” Martín said. “These devices can indicate to the driver approximately where he is, but they cannot be used as a source of information in an Intelligent Transport System like those we have cited.”

The basic elements that make up this system are a GPS and a low-cost inertial measurement unit (IMU). The latter device integrates three accelerometers and three gyroscopes to measure changes in velocity and maneuvers performed by the vehicle. Then, everything is connected to a computer that has an application that merges the data and corrects the errors in the geographic coordinates. Enrique Martí of UC3M’s GIAA explains, “This software is based on an architecture that uses context information and a powerful algorithm (an unscented Kalman filter) that eliminates the instantaneous deviations caused by the degradation of the signals received by the GPS receiver or the total or partial loss of the satellites.”

The current prototype can be installed in any type of vehicle. It is already working on board the IVVI (Intelligent Vehicle based on Visual Information, pictured above), a car that has become a platform for research and experimentation for professors and students at the university.

The LSI and UC3M researchers working on this “intelligent car” can capture and interpret all of the information available on the road, and that drivers use. To do this, the team is using optical cameras, infrareds and lasers to detect whether drivers are crossing the lines on the road, or whether there are pedestrians in the vehicle’s path, as well as to adapt the speed to the traffic signals and analyze the driver’s level of sleepiness in real time.

Next Steps. The researchers will analyze the possibility of developing a system that makes use of the sensors that are built into smartphones, because intelligent telephones are equipped with more than ten sensors, such as an accelerometer, a gyroscope, a magnetometer, GPS and cameras, in addition to Wi-Fi, Bluetooth or GSM communications.

“We are now starting to work on the integration of this data fusion system into a mobile telephone,” said Enrique Martí, “so that it can integrate all of the measurements that come from its sensors in order to obtain the same result that we have now, but at an even much lower cost, since it is something that almost everyone can carry around in his pocket.”

SPAN tightly couples NovAtel’s precise GNSS technology with highly accurate inertial measurement technology to provide a robust, stable and continuous 3D navigation. The new OEM-ADIS-16488 sensor is designed to be coupled with NovAtel’s OEM6 receivers via the MEMS Interface Card (MIC), providing integrators with a  compact, powerful GNSS/INS engine, NovAtel said.

The OEM-ADIS-16488 features low noise gyros and accelerometers in a small, lightweight form factor.  This IMU enables precision measurements for applications that require low cost, high performance and rugged durability.  Tight-coupling of the two technologies enables continuous robust positioning in difficult environments where satellite signals are unreliable or unavailable for short periods of time.

The OEM-ADIS-16488 is now available for order and immediate shipment.

Spatial FOG contains a dual-frequency RTK GNSS receiver that provides 1-centimeter accuracy positioning and supports all of the current and future satellite navigation systems, including GPS, GLONASS, Galileo and Compass.

A next-generation memory backup system allows Spatial FOG to hot start inertial navigation from its last position in 2 seconds and obtain a GNSS fix in as little as 3 seconds. The memory backup system lasts for the lifetime of the product and will provide backup for 24 hours without power.

Spatial FOG’s internal filter runs at 1,000 Hz, and data can also output at this rate over high speed RS232 or RS422. This allows for control of dynamically unstable platforms, the companies said. Spatial FOG is also highly tolerant to both shock and vibration thanks to the performance of the KVH 1750 IMU design and advanced filtering.

Spatial FOG supports a wide range of peripherals including external GNSS receivers, odometers, DVLs, USBLs and NMEA devices. It also supports both industry-standard NMEA output and a binary protocol. Spatial FOG also is easily integrated into retrofits or new designs, said Advanced Navigation.

Modern machines such as wood harvesters can automatically cut trees and remove branches, but an expert is still needed to plan a thinning and to mark the trees to be felled. The process can be accelerated if the forest ranger can virtually mark trees to be cut, using geographic coordinates instead of colored crosses sprayed on the stems. This requires the robotic wood harvester to be able to locate itself accurately to enable automatic navigation to the next tree for cutting.

Absorption of the GPS signal in the forest canopy leads to poor results, however, with errors up to 50 meters and more. Furthermore, the canopy may cause interruptions and signal loss for several seconds. The performance can be even worse on a moving vehicle, where the signal may even get lost until the vehicle reaches an open area or stops.

Other approaches use differential GPS (DGPS) sensors as their main source of position information. However, our experiments using a high-precision DGPS sensor showed that its accuracy is not even close to sufficient for navigating to a single tree. As the DGPS suffers from the same canopy-related disturbances and shielding, it cannot benefit from its theoretical advantages. In pratice, the DGPS system did not update its position at all when signal reception became too weak.

A different approach was needed. We found it in the framework of the Virtual Forest, more precisely in the semantic modelling of forests, where techniques are being developed to delineate single trees from remote sensing data, such as airborne laser scanner data. Along with the trees and their geo-coordinates, the height and the diameter at breast-height are determined. This data can be used to generate a tree map, which can be used for navigation. The map has a mean error between 0.5 and 1.5 meters, which is still below the mean tree distance of about 2.5 meters.

Visual GPS. The idea of Visual GPS is to bring current developments in the field of robotics into the forest and combine them with information on forest inventory so that the result outperforms other navigation approaches. A matching algorithm is run based on a tree map, generated from remote sensing data, and the tree group, which was detected by one or more laser scanners.

We then implemented a particle filter algorithm, as it enables considering different kinds of distributions. Particles are also called random state samples, and each particle is a hypothesis as to what the true world state might be.

In the initialization, particles are distributed uniformly. An importance weight wt is calculated for each particle, incorporating the measurements as described below. A sampling step rejects particles with a low importance weight and replaces them with new particles, which are distributed according to the previous map. This process is repeated until the particle distribution concentrates at one point, and the particle with the highest weight is returned as the result (see Figure 1).

Figure 1. Particle concentration after resampling; wood harvester at center.

A single tree as a landmark cannot be associated with its corresponding tree in the map. However, patterns of tree positions can be matched. We chose a square area to guarantee even particle distribution and short calculation time. Each particle represents a hypothesis for the position of the vehicle and is tested for its probability to represent that position.

To make the approach more robust against faulty tree maps, we implemented a rotation variant approach, determining vehicle heading along with its position. This enhanced the probability measure used in the propagation step. Instead of embracing only the distances of the trees to the reference point, their relative position is used, considering the heading wt of the current particle:

This approach directly calculates vehicle heading, but the sensitivity towards rotation, which results from the new probability measure, leads to a higher number of particles that must be used during the initialization step.

Global Search. Experiments on a test area with about 22,700 trees proved that the algorithm worked reliably for tree groups containing 20 or more trees, and for position errors of the magnitude of the mean tree distance. Similar tree groups could not be found within the forest. However, the calculation time was too long to be used for navigation.

Local Search. To overcome the high calculation time, we reduced the number of particles. The initial position is estimated with an ordinary GPS sensor. Although the GPS measurement is faulty in the forest, it can limit the search to a restricted area. Machines most often start at the edge of a forest stand, at a forest road, or a canopy opening. At these spots the canopy usually is transparent, and GPS sensors work with higher precision. Therefore, they provide a good initialization for the algorithm.

Robotic wood harvester.

In the following steps, the previous position can be used instead of the output of the GPS sensor for determining the search area. The previous position provides a better initial pose estimation than the GPS sensor and therefore gives the opportunity to further decrease the search area.

To reduce the number of trees for which the distance has to be calculated, trees with a distance from the initial pose estimation smaller than the sum of the estimation of the maximal position error and the maximal distance of the trees in the scanned tree group from the reference position are extracted from the tree map.

Another way to reduce the search area is to estimate vehicle orientation. This is difficult for machines such as wood harvester, which moves slowly and stops frequently when cutting trees. Therefore, small lateral position differences result in large orientation deviances, as the difference vector does not directly point into the direction of the movement any more. Another approach is to use sensor fusion and mount a compass onto the vehicle. During particle initialization, the angle can be restricted to the domain of uncertainty around the compass orientation. However, mounting a compass onto a wood harvester proved to be a serious problem, as the harvester’s massive metal body disturbs the compass measurement.

Figure 2 shows the workflow of the complete system.

Figure 2. Navigation system components.

Results

The simple criterion presented here proved to be reliable in the vast majority of cases. Problems can occur when the tree group contains trees that are not part of the tree map (false positive). This can happen due to missing trees in the tree map or faulty tree cognition in the local laser scanner measurement. In the first case, the understory might not have been detected in the airborne laser scanner data. In the second case, other objects like the harvester’s aggregate might have been mistaken for a tree.

The case of trees not detected in the local laser scanner measurements but contained in the tree map (false negative) does not create problems in the pose estimation step. The algorithm searches for a corresponding tree for each unit in the tree group. For a false positive, no corresponding tree can be found, whereas a false negative is simply not considered. However, if the size of the tree group is too small, the estimation errors grow. The minimum number of trees depends on the search area radius. A size of 20 trees proved to generate reliable pose estimations even during the global search. Dropping below 15 trees, the number of faulty position increases rapidly as more similar patterns can be found.

Single faulty positions can be filtered with respect to the movement constraints of a harvester. The velocity is very low, and the orientation cannot jump. In the experiments, cycle times of about 0.5 seconds were reached on a standard PC. As forest machines do not demand very short calculation time, the algorithm proved to run fast enough to allow identification of single felled trees onboard real machines. One application of the algorithm was to support a navigation assistant to the next tree, similar to navigation systems in cars.

To evaluate system accuracy on a real wood harvester, a surveyor’s office was instructed to measure the vehicle’s position at seven distinct locations. At each position, the sensor input data was written to file for several seconds. This data was evaluated, and for each location more than 45 pose estimations were calculated. The mean value of the position error amounted to approximately 0.55 meters.

Future Work

Reliability can be enhanced by using a detailed digital ground model and the cabin tilt in order to detect the area where the laser beams hit the ground, and therefore avoid the detection of false positives. Similarly, the position of the aggregate, which can be measured by integrating sensors in the hydraulic cylinders of the crane, can be cut from the laser scanner measurements and ignored during tree detection, further reducing the amount of false positives in the tree group. With the integration of an outlier rejection step for false positives in the detected tree groups that ignores trees for which no corresponding candidate tree can be found, a more accurate importance factor can be calculated.

Another task is the integration of the algorithm with a Kalman filter to allow real-time performance of the algorithm. Therefore, the Kalman filter is initialized with the pose estimation of the particle filter algorithm, which is also used for continuous checks of the current position estimate, thereby combining two algorithms with different advantages. The Kalman filter allows real-time execution and therefore speeds up the overall navigation algorithm. The particle filter algorithm can periodically check the position estimated by the Kalman filter and correct it. Furthermore, it provides a strong method to cope with two main problems in mobile robotics: the data association problem and the kidnapped robot problem.

Simultaneously, a mapping and map-correction algorithm could be integrated into the system so that understory trees, which cannot be detected using remote sensing data, and deciduous trees, which are more difficult to delineate in airborne laser scanner data, can be added to the tree map.

Jürgen Rossmann is head of the Institute of Man-Machine Interaction at the RWTH Aachen University, where Petra Krahwinkler and Markus Emde are research scientists.