Method for creating a correction function for improving the accuracy of a GPS device

Abstract

A method for creating a correction function to improve the accuracy of a GPS device includes collecting multiple time samples at multiple known locations, where each time sample consists of GPS coordinates and associated satellite data from multiple satellites. The satellite data includes or allows for determination of (i) satellite orientation and altitude of the associated satellite, (ii) signal-to-noise ratio of received signals from the associated satellite, and optionally (iii) pseudoranges. For each time sample, a respective error between the known location and the corresponding GPS coordinate is calculated, and an error correction function is created as a function of the respective GPS coordinates and satellite data by applying deep learning / machine learning techniques to the multiple time samples.

Description

[Technical field]

[0001] The present invention relates to a navigation system and method, particularly for use by pedestrians. [Background technology]

[0002] With the advent of smartphones, GPS has become widely used by pedestrians and drivers as the primary method of navigation. Very importantly, smartphone GPS is used by pedestrians in urban areas instead of traditional paper maps.

[0003] However, the use of GPS by pedestrians is inconvenient, difficult, and sometimes impossible, mainly in urban areas, due to the low positioning accuracy of mobile phone GPS systems, and because of the low accuracy of mobile phone magnetometers, pedestrians do not have the orientation data required for navigation.

[0004] The main reason for the low accuracy of GPS systems in urban areas is that tall buildings block the line of sight to the satellites. Moreover, signals from satellites that do not have a direct line of sight are instead reflected by tall buildings and received by pedestrians' GPS and misinterpreted, resulting in large errors in positioning readings. This phenomenon is called "multiple path reflection."

[0005] Multipath reflections affect motorized vehicles as well as pedestrians, but it is less of a problem for motorized vehicles since they are constantly moving significantly faster than pedestrians and therefore their respective satellite visibility is constantly changing rapidly and errors can therefore be smoothed out. Also, because the location of cars and other road vehicles is constrained to the road, techniques such as "snap to map" and the use of IMU (inertial measurement unit) data, and other information, are used. Additionally, orientation is known with high accuracy from GPS motion and speed.

[0006] This shortcoming is a major issue for companies like Uber and Lyft, whose drivers' successful pickup of passengers is highly dependent on the accuracy of the location reported by the passenger's GPS: Drivers and clients may not even be able to meet, as the client may simply be across the street or across a junction.

[0007] Attempts have been made to solve this problem, including 3D modeling of surrounding buildings, topological analysis, and statistical techniques to ignore satellites with multiple reflections, but to the best of our knowledge, no practical and cost-effective solution has been proposed.

[0008] A wealth of prior art relates to the localization of autonomous vehicles and the optimization of localization strategies used by autonomous vehicles, which are based on driving context, such as the geographic region in which the autonomous vehicle is traveling, the time of day, and the speed of the autonomous vehicle. Conventional systems may use a map database to snap an initial position calculated from navigation satellite system data to physical geographic objects, such as roads, with the final output displayed by a navigation device. By such means, as disclosed, for example, in US20110257885, the rough position provided by GPS satellites may be significantly improved and sufficient to guide the vehicle, whether autonomous or driven by a human.

[0009] US20170307761 discloses a method for collaborative determination of positioning errors in satellite-based navigation systems. Even less precise positioning receivers, such as smartphones, located in a geographical zone where the precise position is unknown can contribute to the generation of precise error corrections in the atmosphere if the number of receivers is sufficiently large.

[0010] US20180124572 relates to correcting GPS-based position information using local correction information in a network of moving objects. If two receivers share the same set of visible satellites (i.e., they process signals received from the same set of satellites in arriving at a positioning solution), the two receivers should experience similar positioning errors. The set of satellites used by a GNSS / GPS receiver in calculating geographic positioning information is dynamic and constantly changing due to many factors.

[0011] US20190147610 in the name of Uber Technologies, Inc. discloses a system and method for detecting and tracking objects based on sensor data received from one or more sensors, where the sensor data is fed to one or more machine learning models including one or more first neural networks configured to detect one or more objects based at least in part on the sensor data, and one or more second neural networks configured to track the one or more objects across a set of sensor data.

[0012] In summary, the prior art teaches techniques for correcting satellite errors and recognizes that error corrections associated with satellites serving multiple receivers in the same geographic zone are made available to all receivers in that zone. The prior art also discloses collating data from a large number of vehicles traversing a geographic area and using the data received therefrom to improve maps that allow autonomous vehicles to safely navigate along the same roads. The prior art also addresses the need for driverless taxis to locate stationary passengers and provides techniques for moving pedestrian navigation, recognizing the fact that traditional correction approaches applied to vehicles are not always applicable to pedestrians. The prior art discusses the use of deep learning and neural networks to improve map data.

[0013] However, the art does not appear to suggest the application of deep learning / machine learning techniques to multiple time samples to batch process GPS positioning errors and create correction functions for each of multiple known positions as a function of the respective GPS coordinates and satellite data.

[0014] Nor has it been proposed to use error correction data associated with satellites in a geographic area to apply corresponding or derived corrections to the GPS navigation systems of pedestrians carrying mobile devices such as smartphones. Summary of the Invention

[0015] Therefore, the present invention aims to provide a method that addresses this need.

[0016] This object is achieved according to the invention by a two-part procedure. The first part is a learning method for creating an error correction function to improve the accuracy of GPS devices in an area of ​​interest, while the second part provides a method for improving the accuracy of GPS coordinates received in the area of ​​interest using the already determined error correction function. Within the context of the description and the appended claims, the term "GPS device" is used to refer to any device having a built-in type GPS module. The GPS device is most commonly a smartphone, but may be any other suitable device configured for GPS positioning.

[0017] Accordingly, in a first aspect, the present invention provides a method for creating an error correction function for improving accuracy of a GPS device, the method comprising: collecting a plurality of time samples at a plurality of known locations, where each time sample consists of GPS coordinates and associated satellite data from a plurality of satellites, where the satellite data includes or permits a determination as to (i) the satellite orientation and altitude of an associated satellite, and (ii) the signal to noise ratio of a received signal from the associated satellite; calculating, for each time sample, a respective error between the known position and the corresponding GPS coordinate; and applying deep learning / machine learning techniques to the plurality of time samples to create an error correction function as a function of each GPS coordinate and the satellite data.

[0018] In a second aspect, the present invention provides a method for improving the accuracy of GPS coordinates received in a region of interest, in which a correction function is obtained, the method comprising the steps of: The present invention provides a method that includes inputting GPS coordinates and associated satellite data into a correction function to obtain a position with improved accuracy.

[0019] The two procedures are mutually independent. In particular, the second procedure cannot be performed until an error correction function related to the required region of interest is derived, but the second procedure can be performed alone if an error correction function exists. In this case, the two procedures are performed consecutively. However, they can also be performed simultaneously, whereby a learning procedure is performed at the same time as performing the second aspect, which is performed for the region of interest or to improve the error correction function by continuously collecting new data and applying deep learning techniques to continuously expanding data sets.

[0020] In some embodiments, the multiple time samples are obtained by traveling through a wide urban area and collecting time samples from a motor vehicle having an onboard GPS device and an improved precision positioning, each of the time samples consisting of GPS coordinates and associated satellite data obtained by the onboard GPS device and the improved precision positioning.

[0021] Conceptually, the principle of the invention is to repeat the above method for multiple locations over time in the required geographical area to obtain multiple time samples for each location. GPS coordinates are coordinates of a GPS device that are usually obtained by triangulating between three or more satellites and represent a rough location, the invention works to improve the accuracy. GPS data is determined from satellite signals, which include the time information when the signal was transmitted by each satellite. The GPS device records the time information when the signal is received, and the difference between the time of transmission and the time of reception reflects the flight time from the satellite to the GPS device, which, when multiplied by the speed of light, yields the pseudorange between the satellite and the receiver. This is called a "pseudorange" because the GPS device's clock is not very accurate, unlike the highly accurate atomic clocks of the satellites. Thus, at least four satellites are required to determine the rough location value of a GPS device in a known manner. The resulting data is stored in a database and analyzed to derive a function for each coarse location, which is applied to the GPS coordinates and uses the satellite data acquired for the coarse location to correct the GPS coordinates to produce a more precise location. The degree of precision may be improved by orders of magnitude, so that if the coarse location is accurate to within 30 meters, the location corrected by the present invention may be accurate to within 3 meters. It will be appreciated that in practice it is not always possible to map every coordinate in space. However, it is assumed that over time a sufficient number of points in close proximity to one another will be mapped, so that the calculated error correction function is equally applicable to the location of any point in space that is slightly off from the mapped point.

[0022] It will be appreciated that each time sample may consist of different GPS coordinates and different satellite data. This implies that different GPS coordinates may have different associated satellite data that may correlate to the same corrected position. The error correction function is derived to provide the same corrected position regardless of which set of GPS coordinates and associated satellite data is applied.

[0023] The GPS coordinates and associated satellite data, as well as the position errors for a number of points in the area of ​​interest, are stored in a database for subsequent processing. In some embodiments, the database is compiled based on data collected by a navigation system such as Waze™. Such a system in any case collects the rough position, i.e. the associated GPS coordinates, from the vehicle GPS device and communicates the refined position data to the vehicle. Thus, the only additional information required is the satellite data corresponding to each of the satellites, from whose signals the rough position of each is obtained. This information is of course known to the on-board GPS, from which information can be collected and used to compile the database and to derive the correction function of the present invention. In use, the rough position received from the GPS device is then fed into the database, along with the satellite data and the refined position coordinates. This database is used to derive the correction function using deep learning techniques.

[0024] A pedestrian carrying a smartphone with a built-in GPS device can at best only obtain a rough location when the smartphone uses a navigation system such as Waze™. This is because, as described above, the improved accuracy applies to vehicles that are constrained by the navigation system to move along a defined path, and cannot be applied to pedestrians who are free to roam. Furthermore, as explained above, errors caused by reflection of multiple paths in urban areas have a greater impact on pedestrian devices that are stationary or moving slower than motorized vehicles. However, since the correction function is based on and applied to the rough location in accordance with the present invention, the correction function can be equally applied to the GPS device in the pedestrian's smartphone and may derive a more accurate location. Similarly, technologies such as radar, lidar, and other feature augmenters being developed for use by advanced driver assistance systems (ADAS) may be used to determine improved location accuracy. The correction function may be downloaded to the pedestrian's smartphone from a remote server that communicates with the pedestrian's GPS device, or may be stored in the pedestrian's GPS device if there is sufficient memory.

[0025] The error correction function receives as input a coarse position corresponding to a GPS coordinate and a vector of associated satellite data from which the GPS coordinate was derived, and produces as output a true position. [Brief description of the drawings]

[0026] In order to understand the invention and how it may be carried out in practice, embodiments will now be described, by way of non-limiting example only, with reference to the accompanying drawings, in which: [Figure 1] 1 is a pictorial representation of a system according to the present invention. [Diagram 2]1 illustrates a schematic representation of a number of time samples collected by a vehicle over a region of interest randomly over time and stored, which may be used to generate a database in accordance with an embodiment of the present invention. [Diagram 3] FIG. 2 is a flow diagram illustrating the main operations performed by a method according to an embodiment of the present invention for improving passenger / driver encounters. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0027] FIG. 1 is a diagrammatic depiction of a system (10) according to an embodiment of the present invention. The system (10) shows a pedestrian (11) located in an area of ​​interest and holding a smartphone (12) with a built-in GPS module that receives GPS data from at least three satellites (13), (13'), (13'') and calculates a rough location in a known manner using triangulation. Throughout this description and the appended claims, this rough location is referred to as the GPS coordinates of the GPS device. Vehicles (14), (14') driving randomly throughout the area of ​​interest also receive GPS data from the satellites and determine their respective rough locations. However, the vehicles are equipped with navigation systems such as WAZE™, SATNAV™, etc., which refine the GPS coordinates so that the correct or true location of the vehicles is known. For purposes of clarity, "true" and "correct" do not intend that the corrected position be precise in an absolute sense, but rather that it be significantly more accurate than a rough position based on coordinates alone. Anyone who uses a navigation system such as WAZE™ is familiar with being told in advance to turn left at an upcoming junction, and then being told "turn left" as you begin to enter the junction. It is this level of accuracy that makes navigation systems so reliable and user-friendly.

[0028] The vehicle's navigation system may improve accuracy to a rough location based solely on satellite data using aiding data based on precise maps, for example, that are pre-compiled and allow the vehicle's position to be corrected using known techniques such as snap-to-map techniques.

[0029] Similarly, technologies such as radar, lidar, and other augmentations being developed for use by advanced driver assistance systems (ADAS) may be used to determine improved position accuracy.

[0030] Although only two vehicles are shown in the figure, in reality there are thousands of vehicles traveling vehicle routes such as highways, roads, trails and even off-road routes over time, and the locations of these vehicles are accurately mapped and accessible to the vehicle navigation system, either because the map data is preloaded or because the map data is accessible online, typically through the Internet (15). The satellite data received by the vehicle navigation system includes or enables a determination of the satellite orientation and altitude of each associated satellite, as well as the signal-to-noise ratio of the received signal from the associated satellite, and optionally other data such as pseudoranges.

[0031] The invention is based on the collection and storage in a database of time samples of coarse position, satellite data, and true position, and derives an error correction function that is collated over time using deep learning from vehicular traffic moving on dedicated routes in an area of ​​interest, where the GPS device is located. To this end, the invention includes two distinct stages, which are described here.

[0032] Data collection and learning: Here, the main operations performed by the first step are described: GPS errors at a particular known location under various satellite and environmental conditions are collected over time, covering a broad set of positioning errors and satellite data.

[0033] While data could theoretically be collected manually by mapping locations in much the same way that military survey maps are compiled, data could also be collected from a motorized vehicle moving within a desired area for a period of time to quickly and automatically obtain sufficient data over a wide area. FIG. 2 shows a schematic of multiple time samples taken randomly by a motorized vehicle passing known locations in an area of ​​interest, where multiple time samples are collected for each known location over an extended period of time. Thus, by way of simple illustration, a vehicle is shown passing randomly through locations x1, x2, and x3, communicating rough GPS location, satellite data, and refined location to a data server. Over a given period of time, v 11 …v 61 Six vehicles specified in pass position x1 and v 12 …v 72 Seven vehicles designated by pass position x2, and v 13 …v 535 vehicles designated as x1, x2, and x3 are seen passing through location x3. Although not critical, note from FIG. 2 that the wall-clock time of each sample is of course different for each known location. The vehicle's GPS navigation system continually receives the rough GPS position and satellite data and derives an improved position. Also, the vehicles have individual designations, but this is for the sake of clarity. In reality, any given vehicle will pass through multiple locations, and therefore not all of the vehicles arriving at the three distinct locations x1, x2, and x3 will be different vehicles. Certainly, it may occur that a vehicle passes through the same known point more than once, for example, on a return trip or repeated trips by the same vehicle at a later time period. The information collected includes the rough GPS position, satellite status data, and true position. This allows positioning errors to be determined. Today, most motorized vehicles have GPS navigation systems, and most motorized vehicles cover urban areas, so data from a large area of ​​interest can be collected over time. As a result, data corresponding to a vast number of locations in a large urban area is collected, each including a rough position and a corrected position, a positioning error, and associated satellite data. These data serve as inputs for deriving an error correction function using deep learning techniques. The error correction function is configured to receive as input a vector of rough positions corresponding to GPS coordinates and associated satellite data, and to generate as output a true position. As more vehicles cross known points in the area of ​​interest, their data may be collected and stored, thus allowing the error correction function to be further refined over time.

[0034] Estimation of GPS deviation using deep neural networks (DNNs).

[0035] Suppose that for a given point with known position (x, y, z), we obtain an estimated position (x', y', z') generated by the GPS system. The difference (x-x', y-y', z-z') between the true position and the GPS generated position is called the GPS deviation. The task we consider is to estimate the GPS deviation in order to correct the position generated by the GPS system and obtain the correct position as accurately as possible.

[0036] The task can be tackled using standard methods for training deep neural networks (DNNs). An example of training a DNN for the task is based on He, K, Zhang, X, Ren, S, Sun, J. 2015 “Deep Residual Learning for Image Recognition” arXiv:1512.03385 (He et al.) and proceeds as follows: The input to the network is a vector of parameters generated by the GPS system (GPS data), including satellite positions, signal strength, time of measurement, and other possible parameters. For each given location, such data is collected at multiple times. The DNN output is an estimated deviation, expressed for example as a probability distribution over the set of deviations. Training is performed using a loss function, such as the cross entropy between the known deviations and the estimated deviations generated by the DNN. A possible state-of-the-art architecture for the task could be a ResNet DNN of sufficient depth as described by He et al., using a standard training procedure, e.g., with respect to splitting data into training and test sets, searching for good training rates, using batch normalization, etc. To obtain a sufficient amount of data for training the network, the training procedure may use data from multiple different locations at multiple times. It is important to note that while the method described by He et al. was applied to images, the general architecture is equally applicable to the present invention.

[0037] Fix: The correction model is used to correct the GPS error at any point as a function of the satellite data (as defined above). Thus, in use, a navigation system or application in a GPS device, such as a smartphone, within an area of ​​interest receives a coarse GPS position and associated satellite data, and applies the correction function to the coarse GPS coordinates to derive a corrected position.

[0038] application Deriving a person's orientation using a handheld digital map device

[0039] While digital maps have greatly simplified navigation for pedestrian use, the need for pedestrians to correctly orient the map remains a frequent drawback. With traditional printed maps, a user identifies where they are located on the map, identifies landmarks such as streets, and then aligns the landmarks on the map with the landmarks in the real world so that the map is correctly aligned with the real world. Smartphones attempt to do this using an inertial magnetic unit (IMU), which attempts to align an internal magnetometer, acting as a compass, with true north so that a user standing at a known location will know whether to turn left or right, or move forward or backward.

[0040] In fact, as tourists attempting to navigate to unfamiliar locations know, this does not work reliably, and it is common for a user to walk for a while and then encounter a street or other landmark only to realize that they have been going in the wrong direction, and thus the street the map shows is in the opposite direction. This also applies to navigation software that provides voice directions.

[0041] In one application, the invention uses a smartphone with a built-in camera to capture a scene with identifiable landmarks and point the smartphone camera at the landmarks. Once properly aligned, an application in the smartphone is manually launched by the user to find the landmarks in the Street View image. This allows the smartphone application to determine the location of the landmarks, and because the user's position corresponding to the smartphone position is precisely known, the application can determine the orientation of the landmarks relative to the user's position.

[0042] Thus, if the user device is a smartphone with a built-in GPS device, a software application for determining orientation may be stored in the smartphone, and the determination of orientation may include: i. determining a location of the smartphone using the GPS device; ii. Aiming an eye gaze unit with an IMU at a distinct object in a real scene; iii. identifying an image of an object on a Street View image displayed on a screen of the smartphone; iv. Extracting object locations from the Street View database; v. calculating the orientation of the object relative to the position of the smartphone so that a user holding the smartphone can orient himself or herself in space relative to the object.

[0043] It should be noted that "Street View" is the name of a proprietary program of Google Inc. Due to the fact that Google Inc. has invested heavily in the development of Street View and its ready availability, the present invention preferably uses Google's Street View. However, any digital map database in which landmarks are resolved to street level may be used, and therefore, within the context of the appended claims, the term "Street View" is not intended to limit the scope of the claims to Google Street View maps, but rather any digital map database with similar functionality.

[0044] Improvements to help drivers and passengers meet better

[0045] In one embodiment, the present invention can be used to enable passengers waiting on street corners or other locations to communicate their exact location to taxi drivers equipped with improved positioning-enabled navigation systems, so that the taxi driver knows exactly where to meet the passenger. Typically, passengers use their mobile devices to obtain their location, which is then communicated to the taxi driver. However, for reasons detailed earlier, the location obtained by the passenger's mobile device is prone to errors. As a result, the taxi driver may not identify the passenger even if the taxi arrives very close to the passenger. In practice, this is usually resolved by the driver calling the passenger, or vice versa, and providing appropriate instructions. However, this is inconvenient, time-consuming, and also infeasible when the taxi is an autonomous vehicle.

[0046] The present invention provides a solution because once a taxi is within distance to pick up a passenger, the GPS positioning error corresponding to the taxi's location will be available to the passenger and can therefore be considered to be used to correct the passenger's GPS positioning. The taxi driver will then receive a rough GPS location from the passenger and his satellite data, and if the satellite data received from the passenger matches the taxi's data, it will apply the corrections so that the passenger's exact location can be determined without the need for verbal instructions.

[0047] More generally, this aspect of the invention provides a method for locating a first entity by a second entity, the first entity carrying a first GPS device and the second entity carrying a second GPS device, the second entity having a navigation system that provides an improved accuracy of the location of the second entity based on the GPS coordinates of the second GPS device. i. receiving respective GPS coordinates of a first GPS device and associated satellite data, where the satellite data includes or permits a determination regarding (i) satellite orientation and altitude of the associated satellite, (ii) signal-to-noise ratio of a received signal from the associated satellite, and (iii) optionally other data such as pseudorange data; ii. comparing the satellite data of the second GPS device when the second GPS device approaches the first GPS device; iii. determining an offset between the GPS coordinates of the second GPS device and the improved precision position of the second entity when the respective satellite data of both GPS devices are substantially coincident; and iv. applying an offset to the GPS coordinates of the first GPS device to provide a better estimate of its correct location; 4 is a flowchart showing the salient operations in a method performed by a second entity, such as

[0048] In particular, it should be noted that features described with reference to one or more embodiments are described by way of example and not as limitations on those embodiments, and thus, any optional feature described with reference to only some embodiments is assumed to be applicable to all other embodiments as well, unless otherwise indicated or unless a particular combination is expressly impermissible.

[0049] It will also be understood that the system according to the invention may be a suitably programmed computer. Likewise, the invention contemplates a computer program readable by a computer for carrying out the method of the invention. The invention further contemplates a machine-readable memory tangibly embodying a program of instructions executable by the machine for carrying out the method of the invention. Typically, the computer is a mobile phone processor, but is not limited to such. It may be any other handheld or head-mounted device. Handheld means a device that is held in the hand in the normal mode of use of the device.

Claims

1. 1. A method for creating a correction function for improving accuracy of a GPS device, the method comprising: collecting a plurality of time samples at a plurality of known locations, where each time sample consists of a GPS coordinate and associated satellite data from a plurality of satellites, where the satellite data includes or permits a determination as to (i) a satellite orientation and altitude of an associated satellite, and (ii) a signal to noise ratio of a received signal from an associated satellite; calculating, for each time sample, a respective error between the known position and the corresponding GPS coordinate; and applying deep learning / machine learning techniques to the plurality of time samples to create an error correction function as a function of each GPS coordinate and the satellite data.

2. 13. A method for improving the accuracy of GPS coordinates of a region of interest, wherein an error correction function is obtained using the method of claim 1, comprising: inputting the GPS coordinates and the associated satellite data into the correction function to obtain a position with improved accuracy.

3. The method of claim 2 , wherein the error correction function is stored on the GPS device.

4. The method of claim 2 , wherein the error correction function is downloaded to the GPS device from a remote server in communication with the GPS device.

5. 5. The method of claim 1, wherein the error correction function is created by collecting time samples over time from a motor vehicle traveling in an extensive urban area and having an on-board GPS device and an improved precision positioning, each of the time samples consisting of the GPS coordinates and associated satellite data obtained by the on-board GPS device and the improved precision positioning.

6. A method according to any preceding claim, wherein the satellite data includes or allows for the determination of (iii) pseudoranges.

7. 1. A method for improving orientation using a handheld smartphone having an integrated GPS device, the method comprising: The GPS device determines the location of the smartphone using the method of claim 2; Aiming a line of sight unit having an IMU at distinct objects in a real scene; Identifying an image of the object on an image of a Street View displayed on a screen of the smartphone; extracting the object's location from the Street View database; and calculating an orientation of the object relative to a position of the smartphone so that a user holding the smartphone can orient himself or herself in space relative to the object.

8. A computer program product having a memory for storing program instructions, which, when executed on a suitable processor, performs the method according to any one of claims 1 to 7.

Images