Apparatus and method for positioning a device

Abstract

Apparatus and methods are provided for determining a location of a device. Positioning is based on acquiring images having unique location identifying features, either naturally occurring or artificially constructed. These location identifying features are used to generate a multi-dimensional signature. The multi-dimensional signature is processed to achieve a compressed signature having a reduced dimensionality. The location of the device can be retrieved using the compressed signature.

Description

Technical Field

[0001] The following description generally pertains to the field of positioning systems. More specifically, the following description relates to apparatus and methods for positioning devices. Background Technology

[0002] Modern cars, mobile phones, and many other devices are typically equipped with positioning systems. Positioning systems are used to measure the device's location. Location information can be used in many applications. For example, navigation and other map-related applications require knowledge of the device's location in order to generate navigation instructions.

[0003] The most common way to obtain precise location information is by using a satellite-based navigation system. In satellite navigation, a device determines its current location by receiving signals transmitted along the line of sight (LOS) from multiple satellites. The device (such as a mobile phone or vehicle) has an electronic receiver that receives signals that can be used for positioning. Typically, global navigation satellite systems do not require internet access; however, in some cases, internet access can be used to assist in the positioning process.

[0004] As mentioned above, positioning signals are received along the line of sight. Therefore, positioning can only be performed when there is a line of sight to at least four satellites. This limits positioning to many locations outdoors. Furthermore, sometimes buildings, mountains, or other obstacles make at least four satellites invisible, thus preventing positioning. Underground positioning, such as in underground parking lots and highway tunnels, or positioning inside buildings, is particularly difficult.

[0005] The visibility problem has been solved by developing indoor positioning systems. Coarse positioning can be achieved using mobile communication networks. Special-purpose indoor positioning systems can provide accurate locations. However, they are typically based on geographic beacons, which transmit signals using Bluetooth, Wi-Fi, or other suitable radio technologies. These beacons are expensive and require maintenance. Another option is to use multiple different sensors and maps, either individually or in combination. For example, maps and odometer readings can be used to support navigation when a vehicle enters a tunnel; however, vehicles often drift rapidly. Furthermore, odometer information is inaccessible when using a mobile phone or tablet instead of an integrated navigation device. Therefore, improvements to positioning systems are needed. Summary of the Invention

[0006] The following disclosure discloses apparatus and methods for determining the location of a device. Location is based on acquiring images with unique location identification features, which are either naturally occurring or artificially constructed. These location identification features are used to generate a multidimensional signature. The multidimensional signature is processed to achieve a compressed signature with reduced dimensionality. The location of the device can be retrieved using the compressed signature.

[0007] In a first aspect, a method for determining the location of a device is disclosed. The method includes: acquiring an image; extracting location identification features from the acquired image; generating a multidimensional signature of the acquired image using the location identification features; generating a compressed signature by reducing the dimensionality of the multidimensional signature of the acquired image; and retrieving the location of the device using the compressed signature. Using location identification features in the acquired image to determine the location of a device or vehicle is advantageous. This enables accurate positioning in locations where satellite positioning is not possible.

[0008] In one implementation of the first aspect, acquiring the image includes acquiring the image in the road direction. Using images acquired in the road direction (whether forward or backward) is advantageous. This helps to accurately measure distances between images and reduces motion blur.

[0009] In one implementation of the first aspect, the method further includes: acquiring a plurality of images; extracting location recognition features from each of the acquired images; generating a multidimensional signature for each of the acquired images using the location recognition features; generating a compressed signature by reducing the dimensionality of each of the multidimensional signatures; and retrieving the location of the device using the plurality of compressed signatures. Using multiple images is advantageous because it can improve the accuracy of the method if one of the images is occluded.

[0010] In one implementation of the first aspect, the method further includes measuring the geographic distance between the first image and the last image. Measuring the distance between the first and last images is beneficial because the distance information improves the accuracy of the method by determining the appropriate length of the multidimensional signature.

[0011] In one implementation of the first aspect, retrieving the location of the device includes: generating a request including the compressed signature; matching the compressed signature of the request with a database including a plurality of compressed signatures, wherein each compressed signature is associated with a location; and determining the location of the device based on the location of the compressed signature that matches the at least one signature. It is advantageous to set up a database in the device so that location can be performed without a network connection, since a network connection may be unavailable when location is needed.

[0012] In one implementation of the first aspect, matching includes finding the most similar compressed signature sequence to the requested compressed signature. Attempting to find the most similar sequence to the requested compressed signature is advantageous. The requested compressed signature may also have position-insensitive features, and an exact match is not required.

[0013] In one implementation of the first aspect, retrieving the location includes: generating a request including the compressed signature; sending the request to an external database suitable for matching the compressed signature of the request with a database including a plurality of compressed signatures associated with the location of the device, based on the location of the compressed signature matching the at least one signature; and receiving the location of the device in response to the request. Using an external database is advantageous, as it may be larger and easier to maintain for frequent updates.

[0014] In one implementation of the first aspect, a machine learning device is used to extract location identification features for finding relevant location identification features. Using a machine learning device in detecting location identification features is advantageous because the machine learning device can be further trained. This improves the localization of all possible locations, not just those of the trained locations.

[0015] In the second aspect, a computer program including computer program code is disclosed. The computer program code, when executed in a computing device, performs the method as described above. It is advantageous to implement the positioning method as software using circuitry for determining location.

[0016] A third aspect discloses an apparatus. The apparatus includes: a camera for acquiring images; and processing circuitry for: extracting location identification features from the acquired images; generating a multidimensional signature of the acquired images using the location identification features; generating a compressed signature by reducing the dimensionality of the multidimensional signature of the acquired images; and retrieving the location of the apparatus using at least one compressed signature including the compressed signature. Using location identification features in acquired images to determine the location of a device or vehicle is advantageous. This enables accurate positioning in locations where satellite positioning is not possible.

[0017] In one implementation of the third aspect, the camera is used to acquire images in the direction of the road. Using images acquired in the direction of the road (whether forward or backward) is advantageous. This helps to accurately measure distances between images and reduces motion blur.

[0018] In one implementation of the third aspect, the camera is used to acquire multiple images; the processing circuit is used to: extract location recognition features from each of the acquired images; generate a multidimensional signature for each of the acquired images using the location recognition features; generate a compressed signature by reducing the dimension of each of the multidimensional signatures; and retrieve the location of the device using the multiple compressed signatures. Using multiple images is advantageous because it can improve the accuracy of the method if one of the images is occluded.

[0019] In one implementation of the third aspect, the processing circuitry is further configured to: measure the geographic distance between the first image and the last image. Measuring the distance between the first and last images is beneficial because the distance information improves the accuracy of the method by determining the appropriate length of the multidimensional signature.

[0020] In one implementation of the third aspect, when retrieving the location of the device, the processing circuitry is further configured to: generate a request including the compressed signature; match one or more compressed signatures of the request with a database including a plurality of compressed signatures, wherein each compressed signature is associated with a location; and determine the location of the device based on the location of the compressed signature that matches the at least one signature. Setting up a database in the device is advantageous because it allows for location tracking without a network connection, which may be unavailable when location tracking is required.

[0021] In one implementation of the third aspect, upon matching, the processing circuitry is further configured to find the most similar compressed signature sequence to the requested compressed signature. Attempting to find the most similar sequence to the requested compressed signature is advantageous. The requested compressed signature may also have position-insensitive features, and an exact match is not required.

[0022] In one implementation of the third aspect, when retrieving the location of the device, the processing circuitry is further configured to: generate a request including the compressed signature; send the request to an external service suitable for matching one or more compressed signatures of the request with a database including multiple compressed signatures associated with the location of the device, based on the location of the compressed signature matching at least one compressed signature; and receive the location of the device in response to the request. Using an external database is advantageous, as it can be larger and easier to maintain for frequent updates.

[0023] In one implementation of the third aspect, the processing circuitry is used to extract the location identification features using a machine learning device for finding relevant location identification features. Using a machine learning device in detecting location identification features is advantageous because the machine learning device can be further trained. This improves the localization of all possible locations, not just the localization of trained locations.

[0024] A fourth aspect discloses a server for location services. The server includes: a database comprising a plurality of compressed signatures associated with a location; and processing circuitry configured to: receive a request including at least one compressed signature; match the at least one compressed signature of the request with the database comprising the plurality of compressed signatures, wherein each compressed signature is associated with a location; and determine the location of the requesting device based on the location of the compressed signature matching the at least one signature. Using an external server to receive requests from a location device is advantageous. The database of the external server may be larger and easier to maintain for frequent updates.

[0025] In one implementation of the fourth aspect, the matching includes finding the most similar compressed signature sequence to the requested compressed signature. It is advantageous to attempt to find the most similar sequence to the requested compressed signature. The requested compressed signature may also have position-insensitive features, and an exact match is not required.

[0026] The above and other objectives are achieved by the subject matter claimed in the independent claims. Other implementations are apparent from the dependent claims, the specification, and the drawings.

[0027] The principles described in this manual can be implemented in hardware and / or software. Attached Figure Description

[0028] Other exemplary embodiments will be described in conjunction with the following figures, wherein:

[0029] Figure 1 An example of a method for generating compressed signatures is shown;

[0030] Figure 2 An example of a scheme for generating compressed signatures is shown;

[0031] Figure 3 An example of the training process for machine learning involving road signature generation is shown;

[0032] Figure 4 An example of a scheme for acquiring vehicles is shown;

[0033] Figure 5 An example of a device for positioning a car is shown.

[0034] In the following figures, the same reference numerals are used to denote the same or at least functionally equivalent features. Detailed Implementation

[0035] The following description is taken in conjunction with the accompanying drawings, which form part of the invention and illustrate specific aspects of the apparatus and methods by way of illustration. It should be understood that other aspects may be utilized and structural or logical modifications may be made without departing from the scope of the claims. Therefore, the following detailed description should not be construed as limiting.

[0036] For example, it should be understood that the disclosure relating to the described method also applies to the corresponding device or system for performing the method, and vice versa. For example, if specific method steps are described, the corresponding device may include units that perform the described method steps, even if such units are not explicitly described or shown in the drawings. Furthermore, it should be understood that features of the various exemplary aspects described herein can be combined with each other unless otherwise explicitly stated.

[0037] Figure 1 Examples of methods for retrieving the location of a device, such as a mobile phone, vehicle, or any similar device including a camera and capable of accessing a database containing location-associated compressed signatures or other similar references that can be derived from images, are disclosed. Figure 1 In this example, the method begins by acquiring an image (step 100). The image is acquired using a conventional camera, such as a mobile phone camera unit or a vehicle camera. The acquired image is then processed to extract location-identifying features from it (step 110). Non-location-identifying features may also be extracted during feature extraction. Other analysis tools can be used to remove non-location-identifying features; however, location can be determined when the acquired image contains sufficient location-identifying features. In some cases, one location-identifying feature is sufficient. For example, some road signs are unique, and when they are extracted, their location can be accurately determined even if the image contains other features. The extraction step may include identifying the extracted features. For example, when a feature is identified as a vehicle, that feature may be discarded because a vehicle is highly unlikely to be used to determine location. A multidimensional signature is generated from the extracted features (step 120). The multidimensional signature provides very accurate identification information for the extracted features. The multidimensional signature is used as a source for generating a compressed signature (step 130). Finally, the location of the device is retrieved using one or more of the generated compressed signatures (step 140).

[0038] An example of a method has been disclosed above. This method can be implemented in a device including processing circuitry. The processing circuitry includes at least one processor for executing a computer program and at least one memory for storing the computer program and related data. Depending on the required task, the memory may be volatile or non-volatile. The processor executes the computer program, thereby performing the method as described above. Two or more devices may be used instead of a single device, as described in the following disclosure.

[0039] Figure 1 The exemplary method can be fully executed within the device. This implementation can be considered an offline implementation. In an online implementation, one or more steps are performed by a remote service such as a cloud service or a location server. Step 100, acquiring an image, is performed within the device being located. Because modern cameras have high resolution, it is generally undesirable to send the acquired image to a remote service over a network. To reduce network traffic, location identification features are typically extracted within the requesting device. Furthermore, generating a multidimensional signature and generating a compressed signature are typically done within the requesting device. This is advantageous because the compressed signature is smaller, so the request sent to the cloud or location service does not require as much bandwidth as the image or multidimensional signature. The location can be retrieved from an internal database stored in the device's memory or from an external service. Retrieval includes generating an appropriate request and receiving a response. Furthermore, retrieval includes processing the request and sending a response. In an offline implementation, these are internal steps and signals within the device. In an online implementation, these tasks are performed by an external service, and the device does not need any further information about these tasks.

[0040] Figure 2 Examples of schemes for generating compressed signatures are disclosed. This scheme can be used in several different devices or apparatuses, such as mobile phones or automobiles. The following disclosure uses an automobile as an example; however, the scheme can be applied to any device with a camera unit for acquiring images and access to a database, which can reside within the device, a network service (e.g., a cloud service), or a server.

[0041] In this scheme, one or more images are first acquired (step 200). In the example of a car, these images are acquired in the direction of the road. This direction can be forward or backward. A single image may be sufficient for coarse localization. However, when using the method provided by this scheme, such as in navigation applications, using multiple images can improve the accuracy of the method. Furthermore, the distance traveled between the first and last images or between each consecutive image can be measured and used for localization. The acquired images can be ordinary images acquired using cameras already integrated into the acquisition device. For example, modern cars and mobile phones typically have one or more cameras.

[0042] The acquired image is sent to the image encoder 210. In the image encoder, the image is processed to find location recognition features in the image. Location recognition features are features that can be used for positioning. Thus, location recognition features are objects, landmarks, etc. Location recognition features do not change or move rapidly over time. Examples of location recognition features include trees, rock formations, lakes, buildings, road signs, roads, and streets, etc. Examples of other features include the sun, people, animals, and other vehicles. These features usually move and change positions, so they cannot be used to determine the location. As a result, the image encoder 210 provides a set of image features including one or more location recognition features 220.

[0043] The image encoder includes a convolutional neural network (CNN), followed by a global pooling layer. The CNN outputs k feature maps, which are aggregated by the global pooling layer into one value (one feature map). Finally, the obtained k-dimensional vector is normalized.

[0044] In Figure 2 the example, the result of the image encoder 210 (i.e., the location recognition features 220) is fed into the signature generator 230. Then, the location recognition features 220 are used together with distance information to generate a signature. The road signature generation module uses a continuous image sequence from time t to t + n and the travel distance d from the position of the camera at time t to the position of the camera at time t + n t+n as input. The number of the generated l vectors is controlled by the spatial resolution factor d Δ of the signature generator, and can be calculated using the following relationship: The attention mechanism inside the signature generator uses a time filter to combine the input image features into road signature vectors.

[0045] The signature generator 230 returns a set 240 of l vectors called road signatures. Each vector of the signature has a dimension k, and the distance between two consecutive vectors is d Δ . Then, the set 240 of l vectors is fed into the dimensionality reduction 250. In Figure 2 the example, dimensionality reduction is an unsupervised machine learning method used to reduce the dimension of the road signature. Dimensionality reduction is applied to each road signature vector to reduce its size from k to m, where m << k. The result is the compressed road signature 260, which is a set of l vectors, where each vector has a dimension m that is much smaller than k, and the distance between two consecutive vectors is d Δ .

[0046] In the image encoding process described above, signature generation and dimensionality reduction are differentiable operations. This means that the signature generation process is a fully differentiable operation, therefore road signatures can be optimized end-to-end for localization purposes, and all road signature generator components can be jointly optimized during training. The training process is described below.

[0047] The process described above is an example of a process that can be used in the examples discussed below. Other similar processes can be used to provide a correlation between one or more acquired images and locations.

[0048] exist Figure 3 The document discloses an example of the machine learning training process involved in the road signature generator. The road signature generator can be trained using the following principles when generating discriminative compressed signatures for precise location.

[0049] Figure 3 The example illustrates two separate pipelines involved in road signature generation. The first pipeline includes receiving a reference image sequence and a distance 300 as reference input, which is processed by a road signature generator 302. This results in the computation of one or more reference compressed signatures 304. The process of generating a road signature can be similar to the one described above. Figure 2 The process is described in the example. The second pipeline receives the target image sequence and distance 310 as target input, which is processed by the road signature generator 304. This results in the calculation of one or more target compressed signatures 314.

[0050] The first pipeline uses a reference input of a continuous sequence of images from time t to t+n, and the travel distance d from the camera's position at time t to its position at time t+n. t+n The second pipeline uses a target input consisting of a continuous sequence of images from time t' to t'+m, and a travel distance d from the camera's position at time t' to its position at time t'+m. t’+m The target sequence is recorded at different times (t≠t') along the same path as the reference sequence. The target sequence is longer than the reference sequence by d. t’+m >d t+n .

[0051] Road signature generators 302 and 312 both use Siamese networks that share the same set of parameter weights to compute reference and target signatures from image sequences. These Siamese networks return compressed signatures for the reference image sequence and compressed signatures for the target image sequence. Because the travel distance in the target sequence is longer than the travel distance in the reference sequence, the target signature is longer than the reference signature.

[0052] The road signature matcher 320 uses a sliding window function, such as zero-mean normalized cross-correlation, to compute the signal similarity 322 between the reference signature and the target signature. The road signature matcher 320 returns similarity measurements 330 of varying sizes based on the signature length. High similarity results in low values ​​in the similarity measurement signal. The road signature matcher 320 creates a ground truth signal based on the spatial and temporal alignment 360 of the ground truth of the reference and target sequences. This is computed by generating a narrow 1D Gaussian function centered at the overlap location of the reference and target sequences. The ground truth alignment of the reference and target sequences is obtained through GPS measurements or content-based image retrieval. A training signal is obtained by computed a cross-entropy loss 350 between the ground truth signal and the similarity measurements obtained from the road signature matcher 320. Before computed loss, the similarity measurements obtained from the road signature matcher 320 are converted into probabilities 340 using a Softmax function 335 on the relative signal. The gradient of the loss is computed based on the input sequence and backpropagated throughout the pipeline to correct the weight parameters of the road signature generator.

[0053] Figure 4 A method for acquiring vehicle 400 for collecting location-related data and constructing a database for image-based localization is disclosed. Figure 4 The vehicle acquisition example includes an acquisition subsystem comprising an inertial navigation system, a satellite-410-based navigation antenna, and an odometer system. The inertial navigation system is coupled to a camera oriented along the vehicle's longitudinal axis. Additionally, the subsystem ensures synchronization between the image acquisition process and the inertial navigation system output. Other subsystems store images in dedicated memory. The acquisition subsystem acquires images with precise geolocation and geotagging along the travel route.

[0054] The road signature generator 430 calculates a compressed road signature based on the recorded image sequence 420 and the distance traveled. Road signature calculation can be performed on an embedded vehicle to reduce in-vehicle storage usage, or on a remote server to reduce the computational burden on the vehicle. The road signature generator 430 generates a compressed road signature 440 with precise longitudinal positioning on the road traveled by the vehicle. The precisely positioned compressed road signature 440 is stored in a geodatabase 450. The geodatabase 450 can then be used as an information source, or the stored information can be packaged for later use.

[0055] Figure 5An example of a device for locating vehicle 500 is shown. The user is in the vehicle, equipped with a device combining a GNSS application and a camera module, with a geographic database stored in its memory. When entering a GNSS-restricted area, the GNSS application triggers an online process that uses an image-based positioning system to determine the vehicle's position. Vehicle 500 acquires images 510 from its camera module, along with the travel distance or vehicle speed (used to derive the travel distance) from the image sequence. The vehicle uses a road signature generator 512 to calculate a compressed road signature 514 from the online-captured image sequence. The compressed road signature 514 calculated online by the user equipment or vehicle is referred to as the target road signature.

[0056] A pre-computed geodatabase of compressed road signatures 520 can be stored, or partially stored, on the user's vehicle or device and queried when entering GNSS-restricted areas. In another implementation, the geodatabase is accessed using the vehicle's internet connection.

[0057] The geodatabase 520 is used to retrieve a reference georeferenced compressed road signature. This is done from the last valid GNSS position received before entering the GNSS restricted area. The user equipment or vehicle retrieves a pre-computed compressed road signature 522 geolocated on the road the user is traveling on from the embedded geodatabase 520. This compressed road signature 522, precisely longitudinally located on the road, is called the reference compressed road signature.

[0058] Vehicle 500 continues to use road signature matcher 530 to match the target signature and the reference signature. Road signature matcher 530 generates a similarity signal between the target signature and the reference signature for similarity measurement 532. The selection of the similarity signal indicates the optimal alignment between the target signature and the reference signature. Therefore, it provides the relative position of the target signature to the reference signature that produces the highest similarity measurement.

[0059] Finally, the user device or vehicle's road signature georeference module 540 uses the similarity signal from the road signature matcher 530 and the precise location of the reference compressed road signature to calculate the precise location 550 of the target compressed road signature along the driving road and derive the user vehicle's current location. Finally, the device location can be displayed in applications that typically rely on satellite positioning.

[0060] As described above, the positioning device can be implemented as hardware such as a mobile phone, tablet computer, computer, telecommunications network base station, or any other network-connected device, or as a method. This method can be implemented as a computer program. The computer program is then executed on the computing device.

[0061] The device, such as a positioning device, is used to perform one of the methods described above. The device includes necessary hardware components. These may include at least one processor, at least one memory, at least one network connection, a bus, etc. For example, the memory or processor may be shared with other components, or accessed from cloud services, centralized computing units, or other resources available via a network connection, instead of dedicated hardware components.

[0062] The apparatus and corresponding methods for positioning have been described in conjunction with various embodiments herein. However, based on a study of the drawings, the invention, and the appended claims, those skilled in the art will be able to understand and implement other variations of the disclosed embodiments in practicing the claimed invention. In the claims, the word "comprising" does not exclude other elements or steps, and "a" or "an" does not exclude multiple. A single processor or other unit may fulfill the functions of several items listed in the claims. The listing of certain measures in dissimilar dependent claims does not imply that combinations of these measures cannot be used advantageously. Computer programs may be stored / distributed on suitable media, such as optical storage media or solid-state media provided together with or as part of other hardware, or may be distributed in other forms, such as via the Internet or other wired or wireless telecommunications systems.

Claims

1. A method for determining the position of a device, characterized in that, The method includes: Multiple images are acquired, wherein the multiple images are a continuous image sequence acquired by the camera in the direction of the road; Extract the location recognition features corresponding to each of the multiple images acquired, the location recognition features being used to characterize the features of objects and / or landmarks that do not change over time or move rapidly; Measure the geographic distance between the first and last images in the plurality of images, or measure the geographic distance between consecutive images in the plurality of images; A signature generator is used to generate a multidimensional signature for each of the plurality of images using the location recognition features and geographic distance corresponding to each of the plurality of images; The signature generator reduces the dimension of the multidimensional signature corresponding to each of the multiple images to generate a compressed signature corresponding to each of the multiple images. The location of the device is retrieved in a database using the compressed signature corresponding to each of the plurality of images, the database including a plurality of compressed signatures associated with the location; The signature generator is trained based on the loss between the similarity measurement between the reference compressed signature and the target compressed signature and the ground truth signal; the reference compressed signature is generated by inputting the distance between the reference image sequence and the corresponding distance into the signature generator to be trained; the target compressed signature is generated by inputting the distance between the target image sequence and the corresponding distance into the signature generator to be trained; the ground truth signal is created based on the spatial and temporal alignment of the ground truth of the reference image sequence and the target image sequence.

2. The method according to claim 1, characterized in that, The road direction can be any of the various directions of a road.

3. The method according to claim 1 or 2, characterized in that, The location of the device to be retrieved includes: A request is made to generate the compressed signature corresponding to each of the plurality of images; Match the compressed signature in the request with a database that includes multiple compressed signatures; The location of the device is determined based on the location of at least one compressed signature that matches the compressed signature in the request among the plurality of compressed signatures.

4. The method according to claim 1 or 2, characterized in that, The location of the device to be retrieved includes: A request is made to generate the compressed signature corresponding to each of the plurality of images; The request is sent to a database containing multiple compressed signatures, and the compressed signatures in the request are matched with the database to obtain the location of the device based on the location of at least one compressed signature among the multiple compressed signatures that matches the compressed signature in the request; The location of the receiving device.

5. The method according to claim 3, characterized in that, Matching involves finding the compressed signature sequence that is most similar to the compressed signature in the request.

6. The method according to claim 4, characterized in that, Matching involves finding the compressed signature sequence that is most similar to the compressed signature in the request.

7. The method according to any one of claims 1, 2, 5 and 6, characterized in that, The location identification features are extracted using a machine learning device for finding relevant location identification features.

8. A device for positioning, characterized in that, include: A camera is used to acquire multiple images, which are a continuous image sequence acquired by the camera in the direction of the road. Processing circuitry, used for: Extract location recognition features corresponding to each of the plurality of images, wherein the location recognition features are used to characterize the features of objects and / or landmarks that do not change over time or move rapidly; Measure the geographic distance between the first and last images in the plurality of images, or measure the geographic distance between consecutive images in the plurality of images; A signature generator is used to generate a multidimensional signature for each of the plurality of images using the location recognition features and geographic distance corresponding to each of the plurality of images; The signature generator reduces the dimension of the multidimensional signature for each of the multiple images to generate a compressed signature for each of the multiple images. The location of the device is retrieved in a database using the compressed signature corresponding to each of the plurality of images, the database including a plurality of compressed signatures associated with the location; The signature generator is trained based on the loss between the similarity measurement between the reference compressed signature and the target compressed signature and the ground truth signal; the reference compressed signature is generated by inputting the distance between the reference image sequence and the corresponding distance into the signature generator to be trained; the target compressed signature is generated by inputting the distance between the target image sequence and the corresponding distance into the signature generator to be trained; the ground truth signal is created based on the spatial and temporal alignment of the ground truth of the reference image sequence and the target image sequence.

9. The apparatus according to claim 8, characterized in that, The road direction can be any of the various directions of a road.

10. The apparatus according to claim 8 or 9, characterized in that, When retrieving the location of the device, the processing circuit is further configured to: A request is made to generate the compressed signature corresponding to each of the plurality of images; The compressed signature in the request is matched against a database containing a plurality of compressed signatures, each of which is associated with a location; The location of the device is determined based on the position of at least one of the plurality of compressed signatures that matches the compressed signature in the request.

11. The apparatus according to claim 8 or 9, characterized in that, When retrieving the location of the device, the processing circuit is further configured to: A request is made to generate the compressed signature corresponding to each of the plurality of images; The request is sent to a database containing multiple compressed signatures, and the compressed signatures in the request are matched with the database to obtain the location of the device based on the location of at least one compressed signature among the multiple compressed signatures that matches the compressed signature in the request; The location of the receiving device.

12. The apparatus according to claim 10, characterized in that, When a match is found, the processing circuitry is also used to find the compressed signature sequence that is most similar to the compressed signature in the request.

13. The apparatus according to claim 11, characterized in that, When a match is found, the processing circuitry is also used to find the compressed signature sequence that is most similar to the compressed signature in the request.

14. The apparatus according to any one of claims 8, 9, 12, and 13, characterized in that, The processing circuit is used to extract the location identification features using a machine learning device for finding relevant location identification features.

15. A server for positioning, characterized in that, include: A database, the database comprising multiple compressed signatures associated with a location; The multiple compressed signatures are generated by the signature generator based on multiple multidimensional signatures. The multiple multidimensional signatures are generated by the signature generator based on multiple location recognition features extracted from multiple images and a first distance. The first distance is the geographical distance between the first image and the last image in the multiple images, or the geographical distance between consecutive images in the multiple images. Processing circuit, the processing circuit being used for: Receive a request that includes at least one compressed signature; Match at least one compressed signature in the request with the database; The location of the device is determined based on the position of the compressed signature that matches at least one of the multiple compressed signatures; Matching includes finding the compressed signature sequence that is most similar to the at least one compressed signature in the request; The signature generator is trained based on the loss between the similarity measurement between the reference compressed signature and the target compressed signature and the ground truth signal; the reference compressed signature is generated by inputting the distance between the reference image sequence and the corresponding distance into the signature generator to be trained; the target compressed signature is generated by inputting the distance between the target image sequence and the corresponding distance into the signature generator to be trained; the ground truth signal is created based on the spatial and temporal alignment of the ground truth of the reference image sequence and the target image sequence.

16. A storage medium, characterized in that, Includes instructions that, when invoked by a processing unit, execute the method of any one of claims 1-7.

Images