System and method for determining the position of a mobile unit of a localization system

Abstract

The invention relates to a method for training (116) an artificial intelligence (110) to determine estimated coordinates (118) of a mobile unit (102) of a localization system, wherein the localization system has a plurality of stationary locating units (106), the mobile unit (102) transmits at least one broadband signal (104), in particular an ultra-broadband signal, at a plurality of training positions, training coordinates (114) of the training positions are determined by means of a training system (112), each broadband signal (104) is received by a plurality of the stationary locating units (106), a respective piece of input information (108) is generated from the broadband signals (104) received by the plurality of stationary locating units (106) for each training position, and the artificial intelligence (110) is trained (116), using the input information (108), to determine the estimated coordinates (118) of each training position of the mobile unit (102).

Description

[0001] System and procedure for determining the position of a mobile unit of a localization system

[0002] The invention relates to a system for determining the position of a mobile unit of a localization system.

[0003] The invention further relates to a method for determining the position of a mobile unit of a localization system.

[0004] A localization system with spatial determination by machine learning is known from WO2023 / 281506A2.

[0005] A method for determining position from a channel impulse response is known from the paper “Single-anchor UWB Localization using Channel Impulse Response Distributions” by Li et al (https: / / arxiv.org / abs / 2211.04246).

[0006] The invention is based on the objective of providing a device and a method mentioned above, which further develop the state of the art.

[0007] This task is solved by the procedures and the system based on independent claims.

[0008] A method for training an artificial intelligence to determine the estimated coordinates of a mobile unit of a localization system is disclosed, wherein the localization system comprises a plurality of stationary positioning units, wherein the mobile unit emits at least one broadband, in particular an ultra-broadband, signal at a plurality of training positions, wherein training coordinates of the training positions are determined by means of a training system, wherein the broadband signal is received by a plurality of the stationary positioning units, wherein input information is generated for each training position from the broadband signals received by the plurality of the stationary positioning units, and wherein the artificial intelligence is trained to determine the estimated coordinates of the respective training position of the mobile unit from the input information.

[0009] A training position is a position of the mobile unit at which it transmits a broadband signal, and whose training coordinates are determined by the training system. The training coordinates are preferably determined at the time the signal is transmitted. The training position, and thus the training coordinates, can also be predefined. The training coordinates serve as a reference for training the artificial intelligence. Each training position has an input value, which is determined from the received broadband signals. The input value can contain a variety of information. For example, it can contain a list of the received signals, with each signal represented as a time-resolved sequence of received values. The signals can also be processed and included in the input value. For example...The corresponding channel impulse response can be calculated for each received signal. In other words, the input information can be represented as a matrix, with each row of the matrix representing a signal or a channel impulse response. Other processing methods for the received signals to obtain input information describing the received signals are also possible. Alternatively or additionally, the input information can include the received signal strength (RSSI) and / or the received angle of attack (AoA). Likewise, additional sensor data acquired by the mobile unit and subsequently encoded in the transmitted signal can be included in the input information. This includes, among other things, barometric pressure information, measurement data from an inertial measurement unit (IMU), or measurements of the Earth's magnetic field to determine spatial orientation.

[0010] The artificial intelligence is trained to output estimation coordinates for each input piece of information, with the estimation coordinates being as close as possible to the training coordinates corresponding to the respective input piece during training. The estimation coordinates are the output of the artificial intelligence, which is optimized during training by changing the AI's parameters. After training, the artificial intelligence outputs estimation coordinates for any input piece of information. Preferably, the training is designed as supervised learning, where each input piece of information and the training coordinates of a training position form a training pair for the artificial intelligence. Thus, at each training position, an input piece of information and the training coordinates of that position are generated.

[0011] Preferably, the input information includes a channel impulse response, in particular one channel impulse response for each received broadband signal. The channel impulse response describes the reaction of a communication channel to a short impulse. It represents the channel's behavior in the time domain and can be used to predict how a signal will be distorted when transmitted through the channel. In a complex environment, the channel impulse response is particularly dependent on the positions of the transmitter and receiver. Existing objects in the environment influence the channel impulse response through reflection and attenuation. Depending on the position of the transmitter and receiver, these physical objects are positioned differently relative to each other and thus affect the channel impulse response accordingly. With a stationary receiver, the characteristics of the channel impulse response allow conclusions to be drawn about the position of the transmitter.With a stationary receiver and a moving transmitter, the channel impulse response changes with the transmitter's movement. With one transmitter and multiple receivers, each receiver has a different channel impulse response. The channel impulse response corresponding to the received UWB signal can be calculated from the UWB signal itself. A preamble in the UWB signal, as defined in IEEE 802.15.4, is particularly suitable for this purpose. The CIR is often expressed as a sequence of values ​​corresponding to different time points. These values ​​are called "taps" and represent the amplitude and phase of the channel response at each time point. A typical channel impulse response consists of a main response, corresponding to the initial arrival of the pulse, followed by several secondary responses. The amplitudes, arrival times, and phases of the responses are described by a statistical distribution.Such a "tap" can be represented as a vector, and the "taps" of all positioning UWB anchors can be combined into a matrix. A matrix created in this way can be used as input information for the artificial intelligence. Preferably, training units with known anchor coordinates receive the broadband signal, and the training coordinates are determined from the propagation times of the broadband signal between the mobile unit and the training units, particularly from propagation time differences. The training units can, for example, be configured as UWB anchors. UWB anchors that serve as positioning units can also be used as training units. The training coordinates can be determined based on the known anchor coordinates using trilateration or triangulation and propagation times and propagation time differences, for example, using the so-called uplink time difference of arrival (UL-TDoA) method.To determine the propagation delay differences, the training units are synchronized and the reception times are compared. To determine the absolute propagation delays, the mobile unit is synchronized with the training units. Preferably, only those training units with a clear line of sight to the mobile unit are used to determine the training coordinates. A clear line of sight between the mobile unit and the training unit can be verified by analyzing the received signal or the channel impulse response.

[0012] Preferably, at least one tracking unit is used as a training unit. The signal received by the tracking unit used as a training unit is thus used both to determine the training coordinates and to generate the input information. This dual use reduces the number of additional training units required.

[0013] Alternatively or additionally, the training coordinates are determined using a simultaneous positioning and mapping (SLAM) system. The SLAM system can, for example, use data from lidar, camera, and / or ultrasonic sensors to determine distances to objects or walls. The sensor data can be compared and / or fused with data from other internal sensors, such as inertial sensors or gyroscopes, and / or external sensors, such as calibrated cameras mounted in the room. This further improves the determination of the training coordinates. Additionally, the SLAM system can operate within the coordinate system of the positioning system using UWB-based self-localization. This eliminates the need for manual referencing of the coordinate systems of both systems. Self-localization can be performed using the downlink TDoA method.In this process, the stationary tracking units emit broadband signals, and the SLAM system determines its own coordinates, and thus the training coordinates, from the time-of-flight differences of the broadband signals and the known coordinates of the tracking units.

[0014] Preferably, a majority of the training positions are selected as reference positions. For retraining the artificial intelligence, the mobile unit emits at least one broadband signal from each of these reference positions. During training, the training coordinates for these reference positions are stored. When retraining the artificial intelligence, a broadband signal is again emitted from the reference position, generating new input information. This new input information, together with the stored training coordinates of the reference position, forms a training pair for retraining the artificial intelligence. By using the reference positions, the training coordinates do not need to be redefined for retraining.

[0015] A further method for determining the estimated coordinates of a mobile unit of a localization system is disclosed, wherein the localization system comprises a plurality of stationary tracking units, the mobile unit transmits a broadband signal, the broadband signal is received by a plurality of the stationary tracking units, input information for a trained artificial intelligence is generated from the broadband signals received by the plurality of stationary units, the artificial intelligence being trained in particular as described above, and the trained artificial intelligence determining the estimated coordinates of the mobile unit from the input information. Through the training of the artificial intelligence, the estimated coordinates essentially correspond to the coordinates that would be obtained through a metrological determination.This determination of the estimation coordinates does not involve any calculation based on runtimes, runtime differences, or other factors.

[0016] No reception angles are necessary. Accordingly, no unobstructed lines of sight are required between the mobile unit and the tracking units. Consequently, significantly fewer stationary tracking units are needed compared to a measurement-based determination of coordinates, as received signals from stationary tracking units without a clear line of sight are also used to determine the estimated coordinates. This contrasts with conventional algorithms, which suppress the influence of multipath propagation signals as much as possible. After the artificial intelligence determines the estimated coordinates, post-processing, such as smoothing or filtering successive estimated coordinates, is possible. Smoothing or filtering can prevent unrealistic jumps in the estimated coordinates. Fusion with parallel conventional algorithms can further improve tracking accuracy.

[0017] Preferably, the input information comprises a channel impulse response, in particular one channel impulse response for each received broadband signal. A channel impulse response contains a wealth of information because the broadband signal is influenced by the surrounding objects. This wealth of information allows the trained artificial intelligence to determine the estimation coordinates more accurately than from signal strength, signal reception time, or other scalar values ​​alone.

[0018] A localization system for determining estimated coordinates of a mobile unit is also disclosed. The localization system comprises a plurality of stationary positioning units and includes a trained artificial intelligence, which is preferably trained according to a method described above. The localization system is configured to perform a method for determining estimated coordinates according to a method described above. The localization system may include a computing unit to utilize the trained artificial intelligence. The positioning units may be configured as UWB anchors. The localization system is preferably configured to perform a method for training an artificial intelligence to determine estimated coordinates according to a method described above.Preferably, the localization system comprises multiple training units for determining training coordinates. These training units can be configured as UWB anchors. The localization units can also be used as training units.

[0019] Preferably, the localization system includes a SLAM system for determining training coordinates.

[0020] Preferably, the artificial intelligence comprises an artificial neural network, in particular a convolutional and / or recurrent neural network.

[0021] A network with recurrent properties is preferred whenever successive position determinations also exhibit a spatial and temporal relationship. This is always the case when the frequency of position determinations is high and the mobile device cannot cover arbitrarily large distances within a single unit of time.

[0022] Preferably, a pre-trained neural network is used, which has been trained independently of a specific positioning system and location. In the actual installation scenario, this network then only needs to be fine-tuned.

[0023] The following description of preferred embodiments, in conjunction with the drawings, serves to further explain the invention. The drawings show:

[0024] Fig. 1 shows a schematic representation of a training procedure;

[0025] Fig. 2 shows an application area for training with training units;

[0026] Fig. 3 shows an application area for training with a SLAM-

[0027] System; and

[0028] Fig. 4 shows an application area for the use of a trained artificial intelligence. Identical or functionally equivalent elements are designated with the same reference numerals in all embodiments.

[0029] Fig. 1 shows a schematic representation of a training method 100. A mobile unit 102 transmits a broadband, preferably an ultra-broadband, electromagnetic signal 104, hereinafter referred to as the UWB signal. The UWB signal is received by a plurality of stationary tracking units 106, here called tracking UWB anchors. From the signals received by the tracking UWB anchors 106, input information 108 for the artificial intelligence 110 to be trained is determined. Preferably, channel impulse responses (CIRs) are used for the input information 108. A channel impulse response typically consists of a main response and further secondary responses generated by reflections of the UWB signal. Particularly preferably, the tracking UWB anchors are temporally synchronized, and the CIRs of the respective tracking UWB anchors have a known temporal relationship to each other.

[0030] A training system 112 determines the training coordinates 114. The training coordinates are the coordinates of the mobile unit at the time the UWB signal 104 is transmitted, i.e., the coordinates of the training position. Examples of determining training coordinates 114 are shown in Figures 2 and 3.

[0031] The mobile unit 102 transmits UWB signals 104 from many different training positions. For each training position from which the mobile unit 102 has transmitted a UWB signal 104, the input information 108 from the received UWB signals 104 and the training coordinates 114 of the training position form a training pair for the training 116 of the artificial intelligence 110. These training pairs enable supervised learning of the artificial intelligence 110. The artificial intelligence 110 is trained to determine estimation coordinates 118 based on the input information 108. During training 116, after comparing the estimation coordinates 118 determined by the artificial intelligence 110 with the training coordinates 114, the artificial intelligence 110 is adjusted so that the artificial intelligence 110 determines estimation coordinates 118 for each input information 108, whereby the estimation coordinates 118 are as close as possible to the training coordinates 114.In the case of an artificial neural network, artificial intelligence 110, adaptation occurs by changing the weights of the artificial neurons within the network. Other artificial intelligence models offer similar adaptation possibilities. After training, the artificial intelligence 110 will output estimated coordinates 118 for each input piece of information 108, even for positions that were not explicitly trained.

[0032] Fig. 2 shows an application area 200 for a training 116 with training units 202 comprising eight rooms 204. In seven of the eight rooms 204, there is one positioning UWB anchor each, serving as a stationary positioning unit 106. In five of the rooms 204, there are additionally two training UWB anchors each, serving as training units 202. A mobile unit 102 is moved along a movement path 206 through the rooms 204, transmitting multiple ultra-wideband signals 104. For clarity, only three positions along the movement path 206 are shown as examples. The ultra-wideband signals 104 are received by the training units 202. In this example, the positioning UWB anchors used as positioning units 106 are also used as training units 202. From the transit times or transit time differences of the ultra-wideband signals 104 to the training units 202 with a clear line of sight, trilateration is used.The training coordinates of the training positions are determined using the uplink TDoA method. Alternatively, the training coordinates can also be determined by triangulation from the receive angle or another method. The positions of the training UWB anchors used as training units 202 must be known to determine the training coordinates. The coordinates of the training units 202 are also referred to as anchor coordinates. The positions of the localization UWB anchors only need to be known if the localization UWB anchors are also used as training units. The training coordinates are later used for training the artificial intelligence. Whether a clear line of sight exists can be determined from the channel impulse response. If no reflections or attenuation are visible in the channel impulse response, but only the main response, it can be assumed that there is a line of sight.From the signals received by the stationary tracking units 106, input information for the artificial intelligence is generated for each training position. In this example, the input information comprises the channel impulse responses (CIRs) of the tracking units 106. Preferably, the tracking UWB anchors are time-synchronized, thus ensuring a known temporal relationship between the channel impulse responses. A large number of training pairs are generated for training the artificial intelligence. Each training pair consists of the training coordinates of a training position and the corresponding input information. The artificial intelligence is then trained to output, for each input piece of information, the estimation coordinates that are as close as possible to the training coordinates of the corresponding training position.

[0033] Fig. 3 shows an application area for training with a SLAM system. The application area is spatially identical to the application area in Fig. 2. Only the differences from Fig. 2 are described below. In this example, no training units are used. The stationary positioning units 106 are in the same positions as in Fig. 2. To determine the training coordinates, a system for simultaneous positioning and mapping, or SLAM system for short, is used in this example. The mobile unit 102 is moved through the rooms 204 by a robot, whereby a map of the rooms 204 is created and the training coordinates are determined as the position of the mobile unit in the room 204 at the time the ultra-wideband signal 104 is emitted. In this example, a lidar scanner is used for this purpose, which scans and maps the walls of the rooms 204 with laser pulses 302.Alternatively, a stereo camera or another mapping system could be used. The input information for each training position is determined in the same way as shown in Fig. 2. The artificial intelligence is trained, as in the example of Fig. 2, with the training coordinates and the corresponding input information. It should be noted that the positions of the positioning UWB anchors in the map do not need to be known for training. Training with a SLAM system is particularly useful for retraining an artificial intelligence, as it requires little preparation. Regular retraining is preferred because physical objects in space can move or change. This prevents the positioning accuracy from decreasing over time due to changes in the space. Decreasing positioning accuracy can be detected using techniques such as action inference, for example, using the gradient descent method.

[0034] Fig. 4 shows an application area for using a trained artificial intelligence. The application area is spatially identical to the application area in Figs. 2 and 3. In this example, the artificial intelligence is fully trained. The mobile unit 102 transmits ultra-wideband signals 104. These ultra-wideband signals 104 are received by the stationary tracking units 106. The input information for the artificial intelligence is generated from the received ultra-wideband signals 104. The artificial intelligence determines the estimated coordinates of the mobile unit 102 from this input information. The position of the mobile unit does not necessarily have to correspond to a trained position. Through training, the artificial intelligence is also able to determine estimated coordinates from the input information for untrained positions.However, the positions of the UWB 106 location anchors must be essentially identical during training and use of the artificial intelligence. Any change in the positions of the UWB 106 location anchors between training and use degrades the accuracy of the estimation of coordinates.

[0035] It is explicitly emphasized that all features disclosed in the description and / or the claims are to be considered separate and independent of one another for the purpose of the original disclosure, irrespective of the combinations of features in the embodiments and / or the claims. It is explicitly stated that all range specifications or specifications of groups of units disclose every possible intermediate value or subgroup of units for the purpose of the original disclosure as well as for the purpose of limiting the claimed invention, in particular also as a boundary of a range specification. List of reference numerals

[0036] 100 training methods

[0037] 102 mobile unit 104 broadband signal

[0038] 106 Location Unit

[0039] 108 Initial Information

[0040] 110 artificial intelligence

[0041] 112 Training system 114 Training coordinates

[0042] 116 Training

[0043] 118 estimated coordinates

[0044] 200 Application area

[0045] 202 Training Unit 204 Room

[0046] 206 Movement path

[0047] 302 laser pulses

Claims

Patent claims 1. Method for training (116) an artificial intelligence (110) to determine estimation coordinates (118) of a mobile unit (102) of a localization system, wherein the localization system comprises a plurality of stationary positioning units (106), wherein the mobile unit (102) emits at least one broadband, in particular an ultra-broadband, signal (104) at each of a plurality of training positions, wherein training coordinates (114) of the training positions are determined by means of a training system (112), wherein the broadband signal (104) is received by a plurality of the stationary positioning units (106), and wherein input information (108) is generated for each training position from the broadband signals (104) received by the plurality of the stationary positioning units (106).where the artificial intelligence (110) is trained (116) to determine the estimated coordinates (118) of the respective training position of the mobile unit from the input information (108) to (102).

2. Method according to claim 1, wherein the training (116) is designed as supervised learning, wherein an input information (108) and the training coordinates (114) of a training position form a training pair for the artificial intelligence (110).

3. Method according to claim 1 or 2, wherein the input information (108) comprises a channel impulse response, in particular a channel impulse response for each received broadband signal (104).

4. Method according to any of the preceding claims, wherein the input information (108) is a received field strength and / or a received angle, in particular a channel impulse response for each received broadband signal (104), includes.

5. Method according to one of the preceding claims, wherein the input information (108) includes information on the mobile unit (102), in particular barometric pressure information, measurement information from an inertial measurement unit (IMU) or a measurement of the Earth's magnetic field to determine the orientation in space.

6. Method according to one of the preceding claims, wherein training units (202) with known anchor coordinates receive the broadband signal (104) and the training coordinates (114) are determined from transit times of the broadband signal (104) between the mobile unit (102) and the training units (202), in particular from transit time differences.

7. Method according to one of the preceding claims, wherein at least one locating unit (106) is used as a training unit (202).

8. Method according to any of the preceding claims, wherein the training coordinates (114) are determined using a simultaneous positioning and mapping (SLAM) system.

9. Method according to one of the preceding claims, wherein a plurality of the training positions are selected as reference positions, wherein, for retraining the artificial intelligence (110), the mobile unit (102) sends out at least one broadband signal (104) from each of the reference positions.

10. Method for determining estimation coordinates (118) of a mobile unit (102) of a localization system, wherein the localization system comprises a plurality of stationary locating units (106), wherein the mobile unit (102) emits a broadband, in particular an ultra-broadband, signal (104), wherein the broadband signal (104) is received by a plurality of the stationary Location units (106) are received, wherein an input information (108) for a trained artificial intelligence (110) is generated from the broadband signals (104) received by the majority of the stationary units (106), wherein the artificial intelligence (110) was trained in particular according to one of claims 1 to 7, wherein the trained artificial intelligence (110) determines the estimation coordinates (118) of the mobile unit (102) from the input information (108).

11. experienced according to claim 10, wherein the estimation coordinates (118) determined by the trained artificial intelligence (110) are filtered and / or fused with information from parallel conventional algorithms.

12. experienced according to claim 10 or 11, wherein the input information (108) comprises a channel impulse response, in particular a channel impulse response for each received broadband signal (104).

13. Method according to claim 12, wherein the input information (108) comprises a plurality of channel impulse responses, wherein a temporal relationship between the channel impulse responses is known.

14. Localization system for determining estimation coordinates (118) of a mobile unit (102), wherein the localization system comprises a plurality of stationary location units (106), wherein the localization system includes a trained artificial intelligence (110), wherein the artificial intelligence (110) was trained in particular according to one of claims 1 to 9, wherein the localization system is configured to perform a method according to one of claims 10 to 12.

15. Localization system according to claim 14, wherein the localization system comprises a plurality of training units (202) for determining training Coordinates (114) included.

16. Localization system according to claim 14 or 15, wherein the localization system comprises a SLAM system for determining training coordinates (114).

17. Localization system according to one of claims 14 to 16, wherein the artificial intelligence (110) comprises an artificial neural network, in particular a convolutional neural network.

18. Localization system according to any one of claims 14 to 16, wherein the artificial intelligence (110) comprises a recurrent neural network, in particular a long short-term memory network.

19. Localization system according to any one of claims 14 to 18, in which the accuracy of the estimation coordinates (118) determined by the artificial intelligence (110) is evaluated according to the principle of action inference.

Images