Improved Navigation and Positioning Using Surface Detection Radar and Deep Learning

Abstract

To provide improved navigation and localization using a surface-penetrating radar and deep learning.SOLUTION: Deep learning to improve or gauge the performance of a surface-penetrating radar (SPR) system for localization or navigation. A vehicle employs a terrain monitoring system including SPR for obtaining SPR signals as the vehicle travels along a route. An on-board computer including a processor and electronically stored instructions, executable by the processor, may analyze acquired SPR images and computationally identify subsurface structures in the acquired SPR images by using the acquired images as input to a predictor. The predictor is computationally trained to identify the subsurface structures in the SPR images.SELECTED DRAWING: Figure 3

Description

Technical Field

[0001] (Cross - Reference to Related Applications) This application claims priority and the benefit of U.S. Provisional Patent Application No. 62 / 900,098, filed on September 13, 2019, the entire disclosure of which is incorporated herein by reference.

[0002] (Field of the Invention) The present invention generally relates to vehicle positioning and navigation, and more generally to improvements in accuracy using deep learning techniques and to system monitoring.

Background Art

[0003] (Background) Surface - detection radar (SPR) systems are used for navigation and vehicle positioning (see, e.g., U.S. Patent No. 8,949,024, the entire disclosure of which is incorporated herein by reference). SPR can be used in environments where GPS accuracy is degraded by multipath or shadowing (such as in cities), or as an alternative to optical sensing approaches (optical sensing approaches cannot tolerate darkness or changing scene illumination, or their performance can be adversely affected by variations in weather conditions).

[0004] Specifically, SPR can be used to obtain scans that include surface features and subsurface features when a vehicle is traversing terrain, and the acquired data scans can be compared to reference scan data previously acquired in the same environment to position the vehicle in that environment. If the reference scan data is labeled with geographical location information, the absolute position of the vehicle can be determined by the labeling.

[0005] Scan data comparison can be, for example, a registration process based on correlation (see, e.g., U.S. Patent No. 8,786,485, the entire disclosure of which is incorporated herein by reference). Positioning of the SPR based on reference scan data overcomes the above limitations of the prior art, but the SPR sensor is not reliable and the alignment process necessarily exhibits some degree of error. For example, the error can result from uncertain ground conditions, aging or malfunction of the SPR sensor, vehicle speed, or variations in ambient conditions (such as wind speed or temperature).

[0006] Accordingly, there is a need for measurements that improve the accuracy of SPR-based positioning systems, minimize the occurrence of error conditions, and estimate the reliability of real-time positioning estimates. SUMMARY OF THE INVENTION MEANS FOR SOLVING THE PROBLEM

[0007] (Overview) Embodiments of the present invention use deep learning to improve or determine the performance of an SPR system for positioning or navigation. The term "deep learning" refers to a machine learning algorithm that uses multiple layers to progressively extract higher-level features from raw images. Deep learning generally involves a neural network that processes information in a manner similar to the human brain. The network consists of a large number of processing elements (neurons) that operate in parallel and are highly interconnected to solve a particular problem. The neural network must be carefully collected and selected training examples and properly trained using the training examples in order to ensure high-level performance, reduce training time, and minimize system bias.

[0008] Convolutional neural networks (CNNs) are often used to classify images or to identify (and classify) objects depicted in an image scene. For example, in the context of self-driving vehicles, CNNs can be employed in a computer vision module to identify traffic signs on the vehicle's route, a person riding a bicycle, or a pedestrian. CNNs use convolutions to extract features from the input image. Convolutions preserve the spatial relationships between pixels and facilitate learning image features using small square input data. Neural networks learn using examples, by which an image can be labeled as containing (or not containing) the feature of interest (autoencoders can learn without labeling). If the system is to be efficiently implemented with confidence, the examples are carefully selected and usually have to be large in number.

[0009] Accordingly, in a first aspect, the present invention relates to a method for detecting and identifying a subsurface structure. In various embodiments, the method includes obtaining an SPR image and computationally identifying a subsurface structure in the obtained image by using the obtained image as an input to a predictor, where the predictor has been computationally trained to identify subsurface structures in SPR images.

[0010] In some embodiments, the method also includes obtaining additional SPR images during traversal of a route, recognizing, by the predictor, subsurface features in the SPR images that the predictor has been trained to recognize, associating the recognized features in the images with corresponding earth coordinates when the image was obtained, and generating an electronic map of the subsurface structure corresponding to the recognized features based on the earth coordinates.

[0011] In various embodiments, the method further includes, while the vehicle is crossing a route, obtaining additional SPR images; recognizing, by a predictor, subsurface features in the SPR images that the predictor is trained to recognize; associating the recognized features with earth coordinates corresponding to the recognized features; and navigating the vehicle based at least in part on the at least partially recognized subsurface features and the earth coordinates of the recognized subsurface features.

[0012] In another aspect, the present invention relates to a system for detecting and identifying subsurface structures. In various embodiments, the system includes an SPR system for obtaining SPR images, and a computer including a processor and electronically stored instructions executable by the processor, the computer analyzing the obtained SPR images and using the obtained images as an input to a predictor to identify, by the computer, subsurface structures in the obtained SPR images, the predictor being trained by the computer to identify subsurface structures in SPR images.

[0013] In any of the foregoing aspects, the predictor can be a neural network (e.g., a convolutional neural network or a regression neural network).

[0014] Yet another aspect of the present invention relates to a vehicle comprising an SPR system for obtaining surface probing radar (SPR) images while the vehicle is in motion, and a computer including a processor and electronically stored instructions executable by the processor, the computer analyzing the obtained SPR images and using the obtained images as an input to a predictor to identify, by the computer, subsurface structures in the obtained SPR images, the predictor being trained by the computer to identify subsurface structures in SPR images.

[0015] In various embodiments, the computer is configured to associate recognized features in an image with corresponding earth coordinates when the image is obtained and, based on the earth coordinates, generate an electronic map of a subsurface structure corresponding to the recognized features. Alternatively, or in addition, the computer can be configured to associate recognized features in an image with corresponding earth coordinates of the recognized features and navigate a vehicle based at least in part on the recognized subsurface features and the earth coordinates of the recognized subsurface features.

[0016] As used herein, the term "substantially" means ±10% by tissue volume and, in some embodiments, ±5% by tissue volume. "Clinically significant" means being considered significant by a clinician, e.g., having an undesired effect on tissue (sometimes, the absence of a desired effect) that triggers the manifestation of damage to the tissue. References throughout this specification to "an example," "one example," "an embodiment," or "one embodiment" mean that a particular feature, structure, or characteristic described in connection with the example is included in at least one example of the technology. Thus, the appearances of the phrases "in one example," "in one example," "an embodiment," or "one embodiment" in various places throughout this specification are not necessarily all referring to the same example. Further, the specific features, structures, routines, steps, or characteristics can be combined in any suitable manner in one or more examples of the technology. The headings provided herein are for convenience only and are not intended to limit or interpret the scope or meaning of the claimed technology. The present invention provides, for example, the following items. (Item 1) A method for detecting and identifying a subsurface structure, the method comprising: obtaining a surface penetrating radar (SPR) image; identifying, by a computer, a subsurface structure in the obtained image by using the obtained image as an input to a predictor, wherein the predictor is trained by a computer to identify a subsurface structure in an SPR image. (Item 2) The method according to Item 1, wherein the predictor is a neural network. (Item 3) The method according to Item 1, wherein the neural network is a convolutional neural network. (Item 4) The method according to Item 1, wherein the neural network is a regression neural network. (Item 5) obtaining an additional SPR image during traversing a route; recognizing, by the predictor, a subsurface feature in the SPR image that the predictor is trained to recognize; associating the recognized feature in the image with corresponding earth coordinates when the image is obtained, and generating an electronic map of the subsurface structure corresponding to the recognized feature based on the earth coordinates. The method according to Item 1, further comprising the above steps. (Item 6) obtaining an additional SPR image during traversing a route by a vehicle; recognizing, by the predictor, a subsurface feature in the SPR image that the predictor is trained to recognize; associating the recognized feature with corresponding earth coordinates; navigating the vehicle based at least in part on the recognized subsurface feature and the earth coordinates of the recognized subsurface feature. The method according to Item 1, further comprising the above steps. (Item 7) A system for detecting and identifying a subsurface structure, the system comprising: an SPR system for obtaining a surface penetrating radar (SPR) image; a computer including a processor and electronically stored instructions executable by the processor. The system is provided with: The computer analyzes the obtained SPR image and identifies, by a computer, a subsurface structure in the obtained SPR image by using the obtained image as an input to a predictor, wherein the predictor is trained by a computer to identify a subsurface structure in an SPR image. (Item 8) The predictor is a neural network, the system according to item 7. (Item 9) The neural network is a convolutional neural network, the system according to item 7. (Item 10) The neural network is a recurrent neural network, the system according to item 7. (Item 11) A vehicle, the vehicle comprising: an SPR system for acquiring a surface perception radar (SPR) image during the running of the vehicle; a computer including a processor and electronically stored instructions executable by the processor; and the computer is configured to analyze the acquired SPR image and use the acquired image as an input to a predictor to identify, by the computer, a subsurface structure in the acquired SPR image, the predictor being trained by the computer to identify the subsurface structure in the SPR image. (Item 12) The computer according to item 11, wherein the computer associates the recognized features in the image with corresponding earth coordinates when the image is obtained, and is configured to generate an electronic map of the subsurface structure corresponding to the recognized features based on the earth coordinates. (Item 13) The computer according to item 11, wherein the computer associates the recognized features in the image with corresponding earth coordinates, and is configured to navigate the vehicle based at least in part on the recognized subsurface features and the earth coordinates of the recognized subsurface features.

Brief Description of the Drawings

[0017] (Brief Description of the Drawings) The above and the following detailed description will be more readily understood when taken in conjunction with the drawings.

[0018]

Figure 1A

[0019]

Figure 1B

[0020]

Figure 2

[0021]

Figure 3

Best Mode for Carrying Out the Invention

[0022] (Detailed Description) Referring first to FIG. 1A, FIG. 1A depicts an exemplary vehicle 102 traveling on a predefined route 104, and the vehicle 102 is provided with a terrain monitoring system 106 for vehicle navigation in accordance with this specification. In various embodiments, the terrain monitoring system 106 includes an SPR navigation and control system 108 having a ground penetrating radar (GPR) antenna array 110 fixed to the front portion (or any suitable portion) of the vehicle 102. The GPR antenna array 110 is generally oriented parallel to the ground surface and extends perpendicular to the direction of travel. In an alternative configuration, the GPR antenna array 110 is closer to or in contact with the surface of the road (FIG. 1B). In one embodiment, the GPR antenna array 110 includes a linear configuration of spatially invariant antenna elements for transmitting GPR signals into the road, and the GPR signals propagate through the road surface into the subsurface region and can be reflected upward. The reflected GPR signals can be detected by the receiving antenna elements within the GPR antenna array 110. In various embodiments, the detected GPR signals are then processed and analyzed to generate one or more SPR images (e.g., GPR images) of the subsurface region along the path of the vehicle 102. If the SPR antenna array 110 is not in contact with the surface, the strongest return signal received can be a reflection caused by the road surface. Accordingly, the SPR image can include surface data (i.e., data regarding the interface between the subsurface region and the atmosphere or local environment). Suitable GPR antenna configurations and systems for processing GPR signals are described, for example, in U.S. Patent No. 8,949,024, the entire disclosure of which is incorporated herein by reference.

[0023] For navigation, the SPR image is compared to a previously acquired and stored SPR reference image with respect to the subsurface area that at least partially overlaps with the subsurface area for a defined route. The image comparison can be, for example, a correlation-based registration process as described in the above-mentioned '485 patent. The location of the vehicle 102 and / or the terrain conditions of the route 104 can then be determined based on the comparison. In some embodiments, the detected GPR signals are combined with other real-time information (such as weather conditions, electro-optical (EO) images, monitoring of the health state of the vehicle using one or more sensors employed in the vehicle 102, and any suitable inputs, etc.) to estimate the terrain conditions of the route 104.

[0024] FIG. 2 depicts an exemplary navigation and control system (e.g., SPR system 108) implemented in the vehicle 102 for navigating driving based on the SPR image. The SPR system 108 can include a user interface 202 through which a user can input data for defining a route or select a predefined route. The SPR image is read from the SPR reference image source 204 according to the route. For example, the SPR reference image source 204 can be a local mass storage device such as a flash drive or a hard disk, or alternatively, or in addition, the SPR reference image source 204 can be cloud-based (i.e., supported and maintained on a web server) and can be remotely accessed based on the current location determined by GPS. For example, the local data store can include SPR reference images corresponding to the vicinity of the vehicle's current location, and when the vehicle is in motion, periodic update information can be read to refresh the data.

[0025] The SPR system 108 also includes a mobile SPR system ("mobile system") 206 having an SPR antenna array 110. The transmission operation of the mobile SPR system 206 is controlled by a control device (e.g., a processor) 208, which also receives the return SPR signal detected by the SPR antenna array 110. The control device 208 generates an SPR image of the subsurface region below the road surface directly beneath the SPR antenna array 110 and / or an SPR image of the road surface.

[0026] The SPR image includes features representing structures and objects (e.g., rocks, roots, boulders, pipes, voids, and soil layers) in the subsurface region and / or on the road surface, and other features indicating variations in soil or material properties in the subsurface / surface region. In various embodiments, a registration module 210 compares the SPR image provided by the control device 208 with the SPR image read from the SPR reference image source 204 to identify the position of the vehicle 102 (e.g., by determining the offset of the vehicle relative to the closest point on the route). In various embodiments, the location information (e.g., offset data or position error data) determined in the registration process is provided to a conversion module 212 that creates a location map for navigating the vehicle 102. For example, the conversion module 212 may generate GPS data corrected for the vehicle's position deviation from the route.

[0027] Alternatively, in one embodiment where the conversion module 212 can read an existing map from a map source 214 (such as another navigation system like GPS or a mapping service) and then position the obtained location information on the existing map, the location map of the default route is stored in a database 216 within the system memory and / or storage device accessible by the control device 208. Additionally, or alternatively, the location data regarding the vehicle 104 can be used in combination with an existing map (such as a map provided by GOOGLE MAPS) and / or data provided by one or more other sensors or navigation systems (inertial navigation system (INS), GPS system, acoustic navigation and ranging (SONAR) system, LIDAR system, camera, inertial measurement unit (IMU) and auxiliary radar system, one or more vehicle dead reckoning sensors (e.g., based on steering angle and wheel odometry), and / or suspension sensors, etc.) for guiding the vehicle 102. For example, the control device 112 can position the obtained SPR information on an existing map generated using GPS. An approach that utilizes the SPR system for vehicle navigation and positioning is described, for example, in the above-mentioned '024 patent.

[0028] An exemplary architecture integrating deep learning with the SPR navigation and control module 108 is illustrated in FIG. 3. As noted above, the system may include various sensors 310 deployed within an associated vehicle. These include SPR sensors, but may also include sensors for identifying conditions related to the operation of one or more deep learning modules 315 as described below. The sensors 310 may also detect external conditions that can affect system performance and accuracy and trigger mitigation strategies. For example, as described in U.S. Serial No. 16 / 929,437, filed July 15, 2020 (which is hereby incorporated by reference in its entirety), the sensors 310 may detect hazardous terrain conditions (and may also trigger a hazard warning in conjunction with SPR measurements of subsurface features). In response, the system 108 may appropriately update the map database 216 and, in some embodiments, may issue a warning to a local authority. The sensors may also capture conditions related to the reliability of position estimation. For example, additional sensors 317 may sense ambient conditions (e.g., wind speed and / or temperature), and hardware monitoring sensors may sense vehicle performance parameters (e.g., speed) and / or vehicle health parameters (e.g., tire pressure and suspension performance) and / or SPR sensor parameters (e.g., sensor health or other performance metrics). Sensor data may be filtered and conditioned by appropriate hardware and / or software modules 323 as in previous cases in the art.

[0029] A plurality of software subsystems implemented as instructions stored in computer memory 326 are executed by a conventional central processing unit (CPU) 330. The CPU 330 can be specialized for the deep learning functions described below or the control device 208 can also be operated (see FIG. 2). An operating system (e.g., MICROSOFT WINDOWS®, UNIX®, LINUX®, iOS, and ANDROID®) provides low-level system functions such as file management, resource allocation, and routing of messages between hardware devices and software subsystems.

[0030] A filtering and conditioning module 335 appropriate for deep learning submodules such as CNNs is also implemented as a software subsystem. For example, if one of the submodules 315 is a CNN, the SPR image can be preprocessed by resizing it to the input size of the CNN, removing noise, smoothing edges, sharpening, etc. Depending on the selected deep learning submodule(s) 315, the generated output can be post-processed by a post-processing module 338 for positioning and for generating metrics such as health status estimation, warnings, and map update information. Post-processing refers to the operation of formatting the output from the deep learning submodule(s) 315 and adapting it to a format and values usable for positioning estimation. Post-processing can include any statistical analysis required to merge the individual deep learning outputs into a single stream of values usable by downstream processing (e.g., averaging the outputs of multiple deep learning modules), changing the data type so that the output can be used in the file format, network protocol, and API required by the download processing. Additionally, post-processing can include conventional non-deep learning algorithms required to convert the output of the deep learning module into a position estimate (e.g., converting a probability density function on an SPR image to a probability density function on the geographical location associated with those SPR images).

[0031] More generally, the positioning may involve adjustments to the predicted vehicle position on the map via the map update module 340, or adjustments to the map itself, and the map update module 340 changes the map from the map database 216 based on the positioning estimates generated by the deep learning submodule(s) 315. The health status metric may include the estimated time to repair, the estimated time to failure for both individual components and the overall system, the accuracy estimate, the estimated damage level of the physical sensor components, the estimate of external interference, and similar indicators of system performance, durability, and reliability. The metric may take the form not only of the confidence value of the overall system and / or individual submodules, but also of the estimated accuracy and error distribution, and of the estimated system latency. The generated metric may also facilitate automatic system parameter tuning (e.g., gain adjustment). These operations may be operated by the metric module 345.

[0032] In one embodiment, the deep learning submodule(s) 315 includes a CNN that analyzes incoming SPR images (e.g., periodically sampled from the GPR antenna array 110) and calculates a probability of match for one or more alignment images. Alternatively, the SPR images can be analyzed in a previous manner as described in the '024 patent to locate one or more best match images and associated probabilities of match. The probability of match can be received from sensors that capture conditions related to the confidence of the position estimate and adjusted based on the input processed by the deep learning submodule 315. This data can be manipulated by different submodules 315 (e.g., another neural network) and can include, or consist of, ambient conditions (e.g., wind speed and / or temperature), vehicle parameters (e.g., speed), and / or SPR sensor parameters (e.g., the health of the sensor or other performance metrics), i.e., any data related to the confidence of the generated SPR data scan and / or the probability of match to a reference image (and thus, for example, the positioning accuracy expressed as an error estimate). The relationship between the data and the SPR images can be the most complex and relevant features that are difficult to detect and use as a positioning reference, which is why a deep learning system is employed. The deep learning submodule 315 receives sensor data and raw SPR images as inputs and outputs a predicted position or data usable by the position estimation module 347, which can estimate the position using image alignment as described above and as described in the '024 patent.

[0033] The method by which the deep learning submodule is trained depends on its architecture, the input data format, and the goal. Generally, a wide range of input data is collected and georeferenced by ground truth information (i.e., known locations) 350. The positioning error of the system is evaluated by a cost function and backpropagated to tune the system weights. The deep learning system can also be trained to recognize subsurface features whose details or conformation can vary. As a conceptual example, a public conduit can vary not only in diameter but also in its orientation with respect to the SPR sensor. Analytically representing all possible SPR images corresponding to the conduit may be unrealistic or impossible, but it is possible to recognize the conduit in any orientation with high accuracy by training a CNN (or, if discrimination requires sequential analysis of multiple SPR images, a regression neural network or RNN). Thus, the neural network recognizes, catalogs the features it is trained to recognize along the subsurface of the root, and thereby associates these with the latitude / longitude coordinates obtained as described above or uses GPS to generate a map of fixed permanent or semi-permanent subsurface structures prior to road or infrastructure construction.

[0034] Alternatively, feature recognition can be used for navigation purposes. For example, knowing that a pipe of a certain size is located at specific GPS coordinates may be sufficient to fix the position of a moving vehicle based on the general detection of pipes in the vicinity of the known location. If multiple features are detected at a known distance apart, the accuracy can be improved. Subsurface features can also represent hazards or suggest the need for preventive measures to avoid hazards. For example, a subsurface area with a high water content under a road can lead to deep holes. If detected and corrected before freeze-thaw cycles, dangerous road conditions can be avoided and the cost of mitigation can be reduced.

[0035] In some embodiments, the deep learning module(s) 315 is hosted locally on computer equipment within the vehicle and can be updated as needed by a server when more intensive training improves the performance of the neural network. In other embodiments, the latest neural network model can be stored remotely, e.g., on the Internet, and accessed by the vehicle via a wireless connection. Map updates and maintenance can also be performed in the "cloud".

[0036] The deep learning module(s) 315 can be implemented using off-the-shelf libraries without undue experimentation. Caffe, CUDA, PyTorch, Theano, Keras, and TensorFlow are suitable neural network platforms (which can be cloud-based or local to an implemented system, depending on design preference). The input to the neural network can be input values, e.g., a vector of readings of SPR scans and system health state information ("feature" vector).

[0037] The control device 208 implemented in the vehicle may include one or more modules implemented in hardware, software, or a combination of both. For embodiments where the functionality is provided as one or more software programs, the program can be written in any of a number of high-level languages such as PYTHON, FORTRAN, PASCAL, JAVA (registered trademark), C, C++, C#, BASIC, various scripting languages, and / or HTML. Additionally, the software can be implemented in assembly language targeted at a microprocessor resident on the target computer. For example, if the software is configured to operate on an IBM PC or a PC clone, the software can be implemented in Intel 80×86 assembly language. The software can be embodied on a manufactured article including, but not limited to, a floppy (registered trademark) disk, a jump drive, a hard disk, an optical disk, a magnetic tape, a PROM, an EPROM, an EEPROM, a field programmable gate array, or a CD-ROM.

[0038] The CPU 330 that executes commands and instructions can be a general-purpose computer, but can also utilize any of a wide variety of other technologies including a dedicated computer, a microcomputer, a microprocessor, a microcontroller, a peripheral integrated circuit element, a CSIC (customer-specific integrated circuit), an ASIC (application-specific integrated circuit), a logic circuit, a digital signal processor, a programmable logic device such as an FPGA (field programmable gate array), a PLD (programmable logic device), a PLA (programmable logic array), an RFID processor, a smart chip, or any other device or arrangement of devices capable of implementing the steps of the process of the present invention.

[0039] The terms and expressions employed in this specification are used as terms and expressions of description and not of limitation, and there is no intention to exclude any equivalents of the features shown or described or parts thereof in the use of such terms and expressions. In addition, although certain embodiments of the invention have been described, it will be apparent to those skilled in the art that other embodiments using the concepts disclosed herein may be used without departing from the spirit and scope of the invention. Accordingly, the described embodiments should be considered in all respects as illustrative and not restrictive.

[0040] The following are the claims.

Claims

A method for positioning a vehicle and detecting and identifying a subsurface structure while minimizing errors caused by aging or malfunction of a surface detection radar (SPR) sensor, the method comprising: acquiring a plurality of SPR images during traversal of a route; acquiring sensor data of conditions related to the reliability of position estimation, the acquired sensor data including parameters of the SPR sensor; identifying, by a computer, a subsurface structure in the plurality of acquired SPR images by using the plurality of acquired SPR images as an input to a predictor, the predictor being trained by a computer to identify a subsurface structure in a plurality of SPR images; estimating the position of the vehicle by using the plurality of acquired SPR images and the acquired sensor data as an input to a deep learning module, the deep learning module calculating a coincidence probability of the plurality of acquired SPR images with respect to one or more alignment images and adjusting the coincidence probability based on the acquired sensor data; A method comprising. The method according to claim 1, wherein the predictor is a neural network. The method according to claim 2, wherein the neural network is a convolutional neural network. The method according to claim 2, wherein the neural network is a regression neural network. The method according to claim 5, wherein acquiring an additional plurality of SPR images during traversal of the route by the vehicle; recognizing, by the predictor, subsurface features in the additional plurality of SPR images that the predictor is trained to recognize; navigating the vehicle based at least in part on the recognized subsurface features and the geocoordinates associated with the recognized subsurface features; The method according to claim 1, further comprising. The method according to claim 1, wherein the identified subsurface structure in the plurality of acquired SPR images includes surface data representing the structure of the road surface and objects.

7. A system for positioning a vehicle while minimizing errors resulting from aging or malfunction of a surface detection radar (SPR) sensor and for detecting and identifying subsurface structures, the system comprising: an SPR system for acquiring a plurality of SPR images and sensor data of conditions related to the reliability of position estimation, the acquired sensor data including parameters of the SPR sensor; a computer including a processor and electronically stored instructions; comprising; the instructions are: analyzing the acquired plurality of SPR images and identifying, by a computer, subsurface structures in the acquired plurality of SPR images by using the acquired plurality of SPR images as inputs to a predictor, the predictor being trained by a computer to identify subsurface structures in a plurality of SPR images; estimating the position of the vehicle by using the acquired plurality of SPR images and the acquired sensor data as inputs to a deep learning module, the deep learning module calculating a probability of coincidence of the acquired plurality of SPR images with one or more alignment images and adjusting the probability of coincidence based on the acquired sensor data; A system executable by the processor to perform.

8. The system according to claim 7, wherein the predictor is a neural network.

9. The system according to claim 8, wherein the neural network is a convolutional neural network.

10. The system according to claim 8, wherein the neural network is a regression neural network.

11. The system according to claim 7, wherein the identified subsurface structures in the acquired plurality of SPR images include surface data representing the structure of the road surface and objects.

12. A vehicle, the vehicle comprising: an SPR system for acquiring a plurality of surface detection radar (SPR) images and sensor data of conditions related to the reliability of position estimation during travel of the vehicle, the acquired sensor data including parameters of the SPR sensor; a computer including a processor and electronically stored instructions; comprising; the instructions are: Analyzing the plurality of acquired SPR images and using the plurality of acquired SPR images as inputs to a predictor to identify, by a computer, a subsurface structure in the plurality of acquired SPR images, wherein the predictor is trained by a computer to identify a subsurface structure in a plurality of SPR images; Estimating the position of a vehicle by using the plurality of acquired SPR images and the acquired sensor data as inputs to a deep learning module, wherein the deep learning module calculates a coincidence probability of the plurality of acquired SPR images with respect to one or more alignment images and adjusts the coincidence probability based on the acquired sensor data; A vehicle executable by the processor to perform the above. The vehicle according to claim 12, wherein the computer is configured to associate the identified subsurface structure with the corresponding geocoordinates when the image is obtained and generate an electronic map based thereon. The vehicle according to claim 13, wherein the identified subsurface structure in the plurality of acquired SPR images includes surface data representing the structure of a road surface and an object.

Images