Adjusting machine learning model for determining intent of person to pass through door

Abstract

There is provided a method for adjusting an intent model (20) that is a machine learning, ML, model for determining an intent of a person to pass through a door (15). The method is performed by an intent determiner (1). The method comprises: determining (40) that the physical environment outside the door has changed; increasing (42) a training rate of the intent model (20), where the training is based on input features based on image data from an image capture device (11) covering an area near the door (15); and applying (44) the intent model (20) based on the image data from the image capture device (11) to infer when the person exhibits the intent to pass through the door.

Description

Technical Field

[0001] The present disclosure relates to the field of machine learning (ML) models for determining a person's intent to pass through a door, and in particular to adjusting such ML models based on changes in the physical environment near the door. Background Art

[0002] Automatic doors have been used for decades. Automatic doors use a motor to open the door when a person approaches and wants to pass through. The door can be a sliding door, a swing door or a revolving door. Automatic doors have the advantage that people do not need to use handles or muscle strength to open the door. The user experience is convenient and has the additional advantage of minimal risk of spreading contaminants through touching surfaces when passing through the door.

[0003] Traditionally, active or passive infrared sensors have been used to detect when a person demonstrates an intent to walk through a door. Other sensors are also known. However, such intent detection is not optimal. In some cases, the person entering needs to be very close to the door to signal the intent to trigger the opening, resulting in a waiting time for the person entering to wait for the door to open. In other cases, the sensors are too permissive and trigger the opening when people walk past the door but have no intent to enter, which can cause climatized (heated or cooled) air to escape from the building housing the automatic door, resulting in energy loss.

[0004] It is well known that evaluating a person's trajectory can more accurately determine intentions. However, how can such trajectory evaluations be updated to reflect the changing physical environment? Summary of the invention

[0005] One goal is to improve how the ML models used to determine intent to walk through a door adapt to changing physical environments.

[0006] According to a first aspect, a method for adjusting an intent model is provided, the intent model being a machine learning (ML) model for determining a person's intent to pass through a door. The method is performed by an intent determiner. The method comprises: determining that the physical environment outside the door has changed; increasing a training rate of the intent model, wherein the training is based on input features based on image data from an image capture device covering an area near the door; and applying the intent model to infer when a person has shown an intent to pass through the door based on the image data from the image capture device.

[0007] The method may further include triggering the door to open when the intention model infers an intention to pass through the door.

[0008] Increasing the training rate of the intent model may include: discarding a previously loaded intent model; and training a new intent model to replace the previously loaded intent model.

[0009] Increasing the training rate of the intent model may include replacing a previous intent model with a new intent model that better corresponds to the determined physical environment.

[0010] Determining that the physical environment outside the door has changed may include determining that the physical environment outside the door has changed based on image data from an image capture device.

[0011] Training can be based on labels derived from whether a person actually walks through a door or avoids walking through a door.

[0012] Determining that the physical environment outside the door has changed may include: classifying at least one object in the image data; determining the position of each of the at least one object; creating a current object map including at least one object and its corresponding position; and determining that the physical environment outside the door has changed when the current object map is sufficiently different from the corresponding previous object map.

[0013] Determining that the physical environment outside the door has changed may include determining that the change in the physical environment persists for more than a threshold duration.

[0014] According to a second aspect, there is provided an intent determiner for adjusting an intent model, the intent model being a machine learning (ML) model for determining a person's intent to pass through a door. The intent determiner comprises: a processor; and a memory storing instructions, which when executed by the processor cause the intent determiner to: determine that the physical environment outside the door has changed; increase the training rate of the intent model, wherein the training is based on input features, the input features being based on image data from an image capture device covering an area near the door; and applying the intent model to infer when a person expresses an intent to pass through the door based on the image data from the image capture device.

[0015] The intent determiner may also include triggering the door to open when the intent model infers an intent to pass through the door.

[0016] The instructions for increasing the training rate of the intent model may include instructions that, when executed by the processor, cause the intent determiner to: discard a previously loaded intent model; and train a new intent model to replace the previously loaded intent model.

[0017] The instructions for increasing the training rate of the intent model may include instructions that, when executed by the processor, cause the intent determiner to replace a previous intent model with a new intent model that better corresponds to the determined physical environment.

[0018] The instructions for determining that the physical environment outside the door has changed may include instructions that, when executed by the processor, cause the intention determiner to determine that the physical environment outside the door has changed based on image data from the image capture device.

[0019] Training can be based on labels derived from whether a person actually walks through a door or avoids walking through a door.

[0020] Instructions for determining that the physical environment outside the door has changed may include instructions that, when executed by a processor, cause the intent determiner to: classify at least one object in the image data; determine the location of each of the at least one object; create a current object graph including at least one object and its corresponding location; and when the current object graph is sufficiently different from the corresponding previous object graph, determine that the physical environment outside the door has changed.

[0021] The instructions for determining that a physical environment outside a door has changed may include determining that a change in the physical environment persists for more than a threshold duration.

[0022] According to a third aspect, there is provided a computer program for adjusting an intent model, the intent model being a machine learning (ML) model for determining a person's intent to pass through a door. The computer program includes computer program code that, when executed on an intent determiner, causes the intent determiner to: determine that the physical environment outside the door has changed; increase the training rate of the intent model, wherein the training is based on input features based on image data from an image capture device covering an area near the door; and apply the intent model to infer when a person has shown an intent to pass through the door based on the image data from the image capture device.

[0023] According to a fourth aspect, there is provided a computer program product comprising the computer program according to the third aspect and a computer readable device comprising a non-transitory memory storing the computer program.

[0024] In general, unless otherwise expressly defined herein, all terms used in the claims should be interpreted according to their ordinary meaning in the technical field. Unless otherwise expressly stated, all references to "a / an / the element, device, component, means, step, etc." should be openly interpreted as referring to at least one instance of the element, device, component, means, step, etc. Unless expressly stated otherwise, the steps of any method disclosed herein do not have to be performed in the exact order disclosed. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Aspects and embodiments will now be described by way of example with reference to the accompanying drawings, in which:

[0026] Figure 1 is a schematic diagram showing an environment in which the embodiments proposed herein can be applied;

[0027] FIG. 2A to FIG. 2Bis a schematic top view showing how a person's trajectory changes based on a physical obstacle near a door;

[0028] FIG. 3A to FIG. 3C is a flow chart illustrating a method for adjusting an intent model, the intent model being an ML model for determining a person's intent to walk through a door;

[0029] Figure 4 It is shown Figure 1 a schematic diagram of components of an intent determiner; and

[0030] Figure 5 One example of a computer program product comprising computer readable means is shown. DETAILED DESCRIPTION

[0031] Various aspects of the present disclosure will be described more fully below with reference to the accompanying drawings, in which certain embodiments of the present invention are shown. However, these aspects may be embodied in many different forms and should not be construed as limiting; rather, these embodiments are provided by way of example so that the present disclosure is comprehensive and complete and fully conveys the scope of all aspects of the present invention to those skilled in the art. Throughout the specification, similar reference numerals refer to similar elements.

[0032] According to the embodiments presented herein, when a physical object is detected that may change the normal trajectory of a person wanting to go through a door, an intent model, which is an ML model for determining intent, is retrained. The intent model determines intent based on the trajectory of the person detected by an image capture device. The trajectory is determined based on image data from the image capture device. When a changed physical environment is detected (based on the image data), the intent model is adjusted to increase its training rate to adjust faster and improve the accuracy of the intent model.

[0033] Figure 1 1 is a schematic diagram showing an environment in which the embodiments presented herein can be applied, as seen in a perspective view. Access to a physical space 16 is controlled by a door 15, which is optionally selectively unlockable. The door 15 is located between a first physical space 14 and a second physical space 16. The door 15 can be a swing door, a sliding door (single leaf or multiple leaves), a revolving door, etc.

[0034] Optionally, in order to control access to the physical space 16, an electronic lock 12 may be provided by selectively unlocking the door 15. In this case, the electronic lock 12 may be controlled to be in a locked state or in an unlocked state. When the electronic lock 12 is in an unlocked state, the door 15 may be opened, and when the electronic lock 12 is in a locked state, the door 15 cannot be opened. In this way, access to the second physical space 16 may be controlled by the electronic lock 12. The electronic lock 12 may evaluate access to the person 5a to 5c by communicating with an electronic key (not shown) of the person 5a to 5c and / or using biometrics. Alternatively or additionally, a conventional mechanical lock may be used to control access through the door 12.

[0035] In passing Figure 1 In the example shown, there are two persons 5a-5b near the door 15. However, it should be noted that the embodiments presented herein are applicable to any number of persons around the door 15.

[0036] An intention determiner 1 is provided for determining a door operation of a door 15. This is based on an image sequence (i.e. a video sequence) captured by an image capture device 11, wherein the image sequence is provided to the intention determiner 1. The image capture device 11 may be a camera, a radar, an infrared sensor, a time of flight (ToF) sensor, etc. Although Figure 1 Only one image capturing device 11 is shown in the first physical space 14 , but a corresponding image capturing device may also exist in the second physical space 16 .

[0037] The intention determiner 1 comprises an intention model 20, which is an ML model for determining a person's intention to pass through a door based on image data from the image capture device 11. Alternatively, part or all of the intention model 20 is provided by a server connected to the intention determiner 1 via a communication network (e.g., the Internet). The image data indicates a movement trajectory of each user 5a to 5b, which can be used as a basis for the intention model 20 to determine when there is an intention (of at least one person) to open the door. The intention model 20 classifies the image data into at least two possible categories: there is an intention of (at least) one person to pass through the door, or there is no intention of anyone wishing to pass through the door.

[0038] When the intention determiner 1 determines that there is an intention to pass through the door, the intention determiner 1 can signal a door opener (not shown) to open the door 15. Using the intention model 20, the intention determiner 1 achieves a balance between convenient door opening and reduced false openings (and any resulting energy losses).

[0039] However, in some cases, there may be new or recently placed physical objects 6 that pose an obstacle to a person wanting to pass through the door. Object 6 may be, for example, a newly installed bicycle rack, a new sign for a building, or a motor vehicle guardrail. Object 6 may affect the trajectory of a person wanting to pass through the door, which may lead to misclassification of intent or no intent.

[0040] The intent determiner 1 optionally includes an object classification model 21, which is an ML model for classifying the type of any physical object 6 that can be identified based on image data from the image capture device 11. Alternatively, part or all of the object classification model 21 is provided by a server connected to the intent determiner 1 via a communication network (e.g., the Internet).

[0041] The intention determiner 1 may be provided as a standalone device, or when an electronic lock 12 is provided, the intention determiner 1 may be integrated with the electronic lock 12 , in which case the electronic lock 12 is a host device of the intention determiner 1 .

[0042] FIG. 2A to FIG. 2B is a schematic top view showing how the trajectory of a person changes based on a physical obstacle near a door. ...

[0043] exist Figure 2A In the example shown in FIG. 1 , there is a first person 5a and a second person 5b. Both persons 5a-5b are moving to pass through the door 15 but not to enter. The intention determiner will use the image data from the image capture device 11 to classify (ie infer) that there is no intention for anyone to pass through the door.

[0044] exist Figure 2B In the figure, it is shown that Figure 2A A similar situation to the case of Figure 2B In the example, there is an object 6 that is an obstacle to passing through the door 5. In this case, Figure 2B In the example shown in Figure 2B , but in this case the first person 5a may actually show an intention to go through the door 15. For example, if the first user 5a has turned to the left to pass the object 6, this can be seen.

[0045] If used for Figure 2A The intent model of the example is also used Figure 2BFor example, the intention of the first person 5a may be incorrectly classified as not having the intention to pass through the door. Therefore, according to the embodiment proposed in this article, the intention determiner 1 increases the training rate of the intention model 20 to adapt to the new physical environment more quickly.

[0046] FIG. 3A to FIG. 3C 1 is a flow chart showing a method for adjusting an intention model 20, which is an ML model for determining a person's intention to go through a door 15. The intention model may be based, for example, on a lower layer including multi-object tracking (MOT) visual object tracking (VOT), which in turn may be based on a convolutional neural network (CNN) and / or object detection technology. The person's posture (including position and direction) may also be estimated. The intention model may also include a classification layer for classifying the intention according to the lower layer, for example, based on an ML classification (e.g., based on CNN) or a regression model, for example, based on a regression neural network technology. When a regression model is used, the output may be applied to predict the probability that a person wants to go through a door. A probability threshold may then be applied so that when the probability is greater than a (configurable or dynamically determined) probability threshold, the intention is determined to be strong enough to be classified as an intention to go through a door. Alternatively, a regression model may be used to predict how long it will take until someone wants to go through a door. When the predicted time is less than a (configurable or dynamically determined) future threshold duration, this may be classified as an intention to go through a door.

[0047] The method is performed by the intent determiner 1. The method may be performed for each image captured by the image capture device, or once for every n number of images captured by the image capture device, where n is a configurable integer. The image may be part of a sequence of images that may form a video sequence, and may have any suitable frame rate, where a higher frame rate results in higher accuracy, and a lower frame rate results in lower resource (e.g., processor usage, power usage) usage.

[0048] In a determine physical change step 40, the intent determiner 1 determines that the physical environment outside the door has changed. The determination may be based on image data (including historical image data) from the image capture device 11 to detect when a change has occurred. Alternatively or in addition, the determination of the change may be based on people adopting new trajectories to pass through the door and / or past the door. For example, if almost all people avoid an area that they did not avoid before when approaching the door, then there is a high probability that there are objects in that area that constitute an obstacle. The assessment of physical changes may, for example, be based on a regression model. In this way, we can distinguish between small and large changes in the environment and update the intent model training accordingly.

[0049] In one embodiment, determining that the physical environment outside the door has changed includes determining that the change in the physical environment lasts for more than a threshold duration. This prevents temporary transient changes from triggering adaptation of the intent model, such as when a bicycle or service vehicle is temporarily parked outside the door for a short period of time. Optionally, the duration depends on the type of object (e.g., as described in more detail below, classified using the object classification model 21).

[0050] In an increase training rate step 42, the intent determiner 1 increases the training rate of the intent model. The training can be an unsupervised training of the intent model based on observations and derived labels. The training is repeated at a certain rate, where the rate is configurable, for example the rate is increased in this step. When the training rate is increased, the training is more frequent, and when the training rate is reduced, the training is less frequent. The training is based on input features that are based on image data from an image capture device 11 covering an area near the door 15. The training can also be based on labels derived from the detected actions of a person, i.e., based on detecting whether a person actually passes through a door or avoids passing through a door. The training can be performed to train the lower layers. Alternatively or additionally, the training can be performed to train the classification layer. Alternatively, the entire intent model can be trained.

[0051] In one embodiment, the intent determiner 1 replaces the previous intent model with a new intent model that better corresponds to the determined physical environment. For example, in December of each year it may be the case that an object in the form of a Christmas tree is placed in front of an entrance. When a similar object is detected in December, the intent model from December of the previous year may replace the currently loaded model in order to more or less immediately adapt the intent model to the Christmas tree obstacle. Optionally, the intent model of another door of a similar scenario may be used as a replacement intent model.

[0052] By increasing the training rate, more previous intent models are forgotten. This results in the intent model adapting to new physical environments faster than a continuously trained intent model; even if such an intent model eventually produces an intent model adapted to new physical environments, it will take longer because with continuous training, it is not desirable to adjust the model based only on the most recent training.

[0053] In an apply intention model step 44, the intention determiner 1 applies the intention model to infer when a person shows an intention to go through a door based on the image data from the image capture device 11. The image data used as input includes both current image data and historical image data to capture the person trajectory.

[0054] In an optional trigger opening step 46, when the intention model infers an intention to pass through the door, the intention determiner 1 triggers the door 15 to open. This can be performed by sending a command to the door controller of the door to open the door. Optionally, the opening of the door can also be conditional on the access control of the person approaching, for example based on an electronic key and / or biometrics.

[0055] Note that the intent model can then be continually trained, where any erroneous intent determinations can be used as training events. For example, if intent is not (erroneously) determined, and the user walks through a door (which can be opened by explicit user intent using a button or infrared beam), then this is used as a training event to reduce the risk of such erroneous non-intent determinations in the future. Similarly, if intent is (erroneously) determined, and the user does not walk through a door, then this is used as a training event to reduce the risk of such erroneous intent determinations in the future.

[0056] Watch now Figure 3B , Figure 3B An embodiment of how the training rate of the intent model may be increased in the increase training rate step 42 is shown.

[0057] In an optional discard previous model step 42a, the intent determiner 1 discards the previously loaded intent model.

[0058] In an optional train new model step 42b, the intent determiner 1 trains a new intent model to replace the previously loaded intent model. The new intent model may be completely blank or may be a predetermined basic intent model.

[0059] Watch now Figure 3C , Figure 3C An embodiment is shown of how changes in the physical environment may be determined in the determine physical change step 40 .

[0060] In an optional classification step 40a, the intent determiner 1 classifies at least one object in the image data, for example using an object classification model 21, which is an ML model trained to detect and classify different objects based on the image data. The classification model 12 can be implemented, for example, using a convolutional neural network (CNN) as known in the art itself. The detected objects can be different object types, such as bicycle racks, traffic barriers, bicycles, cars, trash cans, road signs, Christmas trees, etc. This step can include: for example, based on a CNN trained to classify objects in the image, classifying each object into a specific object type.

[0061] In an optional determine position step 40b, the intention determiner 1 determines 40b the position of each of the at least one object. Each position may be in a coordinate system of the image and / or mapped to a coordinate system of the real world space, such as coordinates in a floor / ground plane.

[0062] In an optional create object graph step 40c, the intent determiner 1 creates a current object graph comprising at least one object and its corresponding position. An object graph is a data set (eg, a collection of data structures, where each data structure comprises an object class and its position).

[0063] In a determine changes step 40d, the intention determiner 1 determines that the physical environment outside the door has changed when the current object graph is sufficiently different from the corresponding previous object graph (from a previous iteration of the method).

[0064] For example, the introduction or removal of an object (optionally, whose size is greater than a threshold size) may be sufficiently different. Alternatively or additionally, the movement of an object beyond a threshold distance (in the image's coordinate system or in a real-world coordinate system) may be sufficiently different.

[0065] Optionally, when the environment outside the door does not change for a long time (duration longer than a threshold duration), the training rate is reduced so as not to reduce resource consumption.

[0066] Figure 4 It is shown Figure 1 Schematic diagram of the components of the intent determiner 1. It should be noted that when the intent determiner 1 is implemented in a host device, one or more of the components mentioned can be shared with the host device. The processor 60 is provided using any combination of one or more of a suitable central processing unit (CPU), a graphics processing unit (GPU), a multiprocessor, a microcontroller, a digital signal processor (DSP), etc. that can execute software instructions 67 stored in the memory 64, so the memory 64 can be a computer program product. The processor 60 can alternatively be implemented using an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), etc. The processor 60 can be configured to perform the above reference FIG. 3A to FIG. 3C Describe the method.

[0067] The memory 64 may be any combination of random access memory (RAM) and / or read-only memory (ROM). The memory 64 also includes a non-transitory persistent storage device, which may be, for example, any single memory or a combination of magnetic memory, optical memory, solid-state memory, or even remotely mounted memory.

[0068] A data memory 66 is also provided for reading and / or storing data during execution of software instructions in the processor 60. The data memory 66 may be any combination of RAM and / or ROM.

[0069] An I / O interface 62 is provided for communicating with external and / or internal entities using wired communication, for example based on Ethernet or a Universal Serial Bus (USB), and / or wireless communication, for example Wi-Fi and / or a cellular network, the wireless communication being compliant with any one or a combination of the next generation mobile networks (fifth generation, 5G), LTE (Long Term Evolution), UMTS (Universal Mobile Telecommunications System) utilizing W-CDMA (Wideband Code Division Multiple Access), or any other current or future wireless network, as long as the principles described below apply.

[0070] In order not to obscure the concepts presented herein, other components of the intent determiner 1 are omitted.

[0071] Figure 5 An example of a computer program product 90 comprising a computer readable device is shown. On the computer readable device, a computer program 91 may be stored in a non-transitory memory. The computer program may cause a processor to perform a method according to the embodiments described herein. In this example, the computer program product is in the form of a removable solid-state memory such as a Universal Serial Bus (USB) drive. As described above, the computer program product may also be stored in a memory of a device such as a Figure 4 Although the computer program 91 is schematically shown here as part of a removable solid-state memory, the computer program may be stored in any manner suitable for a computer program product, such as another type of removable solid-state memory or an optical disc, such as a CD (Compact Disc), a DVD (Digital Versatile Disc) or a Blu-ray Disc.

[0072] The above mainly describes various aspects of the present disclosure with reference to several embodiments. However, as those skilled in the art will readily appreciate, within the scope of the present invention as defined by the appended patent claims, other embodiments other than the embodiments disclosed above are equally possible. Therefore, although various aspects and embodiments have been disclosed herein, other aspects and embodiments are obvious to those skilled in the art. The various aspects and embodiments disclosed herein are for illustrative purposes and are not intended to be limiting, and the true scope and spirit are indicated by the appended claims.

Claims

1. A method for adjusting an intention model (20), the intention model (20) being a machine learning (ML) model for determining a person's intention to pass through a door (15), the method being performed by an intention determiner (1), the method comprising: Determine (40) that the physical environment outside the door has changed; increasing (42) a training rate of the intent model (20), wherein the training is based on input features based on image data from an image capture device (11) covering an area near the door (15); as well as Based on image data from the image capture device (11), the intention model (20) is applied (44) to infer when a person has indicated an intention to pass through the door.

2. The method according to claim 1, further comprising: When the intention model (20) infers an intention to pass through the door, the door (15) is triggered (46) to open.

3. The method according to claim 1 or 2, wherein: Increasing (42) the training rate of the intent model (20) includes: discarding (42a) a previously loaded intent model (20); and A new intent model (20) is trained (42b) to replace the previously loaded intent model (20).

4. The method according to claim 1 or 2, wherein: Increasing (42) the training rate of the intent model (20) includes replacing a previous intent model (20) with a new intent model (20) that better corresponds to the determined physical environment.

5. A method according to any one of the preceding claims, wherein: Determining (40) that the physical environment outside the door has changed includes: determining that the physical environment outside the door has changed based on image data from the image capture device (11).

6. A method according to any one of the preceding claims, wherein: The training is based on labels derived from whether a person actually walks through the door or avoids walking through the door.

7. A method according to any one of the preceding claims, wherein: Determine (40) that the physical environment outside the door has changed including: classifying at least one object in the image data (40a); determining (40b) a position of each of the at least one object; creating (40c) a current object graph including the at least one object and a corresponding position of the at least one object; and When the current object graph is sufficiently different from the corresponding previous object graph, it is determined (40d) that the physical environment outside the door has changed.

8. A method according to any one of the preceding claims, wherein: Determining (40) that the physical environment outside the door has changed includes determining that the change in the physical environment persists for more than a threshold duration.

9. An intention determiner (1) for adapting an intention model (20), the intention model (20) being a machine learning (ML) model for determining a person's intention to pass through a door (15), the intention determiner (1) comprising: Processor (60); as well as A memory (64) storing instructions (67) which, when executed by the processor, cause the intent determiner (1) to: Make sure the physical environment outside the door has changed; increasing a training rate of the intent model (20), wherein the training is based on input features based on image data from an image capture device (11) covering an area near the door (15); as well as Based on image data from the image capture device (11), the intention model (20) is applied to infer when a person has indicated an intention to pass through the door.

10. The intention determiner (1) according to claim 9, further comprising: When the intention model (20) infers an intention to pass through the door, the door (15) is triggered (46) to open.

11. The intention determiner (1) according to claim 9 or 10, wherein: The instructions for increasing the training rate of the intent model (20) include instructions (67) that, when executed by the processor, cause the intent determiner (1) to: discarding the previously loaded intent model (20); and A new intent model (20) is trained to replace the previously loaded intent model (20).

12. The intention determiner (1) according to claim 9 or 10, wherein: Instructions for increasing the training rate of the intent model (20) include instructions (67) that, when executed by the processor, cause the intent determiner (1) to replace a previous intent model (20) with a new intent model (20) that better corresponds to the determined physical environment.

13. The intention determiner (1) according to any one of claims 9 to 12, wherein: The instructions for determining that the physical environment outside the door has changed include instructions (67) that, when executed by the processor, cause the intention determiner (1) to determine that the physical environment outside the door has changed based on image data from the image capture device (11).

14. The intention determiner (1) according to any one of claims 9 to 13, wherein: The training is based on labels derived from whether a person actually walks through the door or avoids walking through the door.

15. The intention determiner (1) according to any one of claims 9 to 14, wherein: The instructions for determining that the physical environment outside the door has changed include instructions (67) that, when executed by the processor, cause the intention determiner (1) to: classifying at least one object in the image data; determining a position of each of the at least one object; creating a current object graph including the at least one object and a corresponding position of the at least one object; and When the current object graph is sufficiently different from a corresponding previous object graph, it is determined that the physical environment outside the door has changed.

16. The intention determiner (1) according to any one of claims 9 to 15, wherein: The instructions for determining that a physical environment outside a door has changed include determining a change in the physical environment that persists for more than a threshold duration.

17. A computer program (67, 91) for adjusting an intent model (20) being a machine learning (ML) model for determining a person's intent to pass through a door (15), the computer program comprising computer program code which, when executed on an intent determiner (1), causes the intent determiner (1) to: Make sure the physical environment outside the door has changed; Increasing the training rate of the intent model (20), wherein The training is based on input features based on image data from an image capture device (11) covering an area near the door (15); as well as Based on image data from the image capture device (11), the intention model (20) is applied to infer when a person has indicated an intention to pass through the door.

18. A computer program product (64, 90) comprising a computer program according to claim 17 and a computer readable device comprising a non-transitory memory storing the computer program.

Images