A method and system for dispensing drinks

Abstract

A system may comprise an artificial neural network to assess user intoxication and to prevent dispensing an alcoholic drink if so. The neural network does this based on image data, e.g., obtained by (e.g., infrared, thermal, depth) cameras 1,3, relating to the user gait, pupil size, and horizontal gaze nystagmus. A user request for an alcoholic beverage may be denied or accepted by an artificial neural network based on the above input data, and a drink dispenser 10 instructed accordingly. A machine learning algorithm may be used in an intoxication detection system for automated bar environments to correlate anonymised user metrics with intoxication level. The detection system includes a processor to analyse visual and physiological metrics related to inebriation. Ambient sensors may detect lighting levels. Each physiological metric may be processed by a dedicated neural network and their assessments combined.

Description

Field The present application relates to a method and system for dispensing drinks, especially the dispensing of alcoholic drinks. Background Many jurisdictions impose legal restrictions on the sale of alcoholic drinks (beverages). For example, in the United Kingdom, the sale of alcoholic drinks is generally subject to a license. Furthermore, alcoholic drinks cannot be sold to minors (those aged below 18) or to a person who is already drunk. Drink vending (dispensing) machines have been developed to provide self-service (automated) vending and dispensing of drinks. In some cases, the beverage may be supplied in a closed container such as a can for subsequent opening by the customer. In other cases, the beverage may be supplied, for example, in a cup or other form of drinking vessel. There has been a significant increase over recent years in the provision of drink vending machines that automate the supply of alcoholic drinks to customers. As one example, such machines have been utilised at festivals, such as music festivals. These events may experience very high levels of demand for drinks at certain timings, such as during the interval between the sets of different performers. The use of automated, self-service drinks machines (systems) helps festival organisers handle peaks in demand. In addition, the adoption of automated drinks machines may support potential customisation of the user experience. The operation of such automated drink dispensing machines is still subject to legal restrictions (licensing requirements). To this purpose, the festival organisers (or their subcontractors, such as the supplier of the automated drinks machines) generally provide human monitoring of the vending process. Such monitoring may check, for example, that would-be users of the automated drinks machines are not minors or drunk. It is relatively easy for such monitors to confirm that a would-be purchaser is an adult (rather than a minor) by checking suitable ID, such as a driving license. However, the situation is more difficult in terms of deciding whether a would-be purchaser is drunk (and hence not allowed to purchase further alcohol). The most common method of making this decision is based on a visual assessment of the would-be purchaser, however, this is somewhat subjective. In practice, with very large numbers of customers, there will be a certain error rate of (i) non-drunk (sober) people being refused from purchasing alcoholic drinks (‘false positives’), and (ii) drunk people being allowed to purchase alcoholic drinks (‘false negatives’). The situation is slightly different from conventional pubs or bars in which a bartender may have to make a similar determination. A bartender who runs such an establishment is likely to have more experience, credibility and control compared with a monitor in a large-scale festival, especially given the transient nature of such events, and the large number of customers that are subject to oversight by such a monitor. There is the possibility of using additional equipment at a festival to support the determination of whether a would-be purchaser is drunk (intoxicated or inebriated). For example, a physical device such as a breathalyser or similar may be used to acquire a physical / chemical measurement of alcohol levels in a would-be purchaser. However, the use of such additional equipment raises costs and may potentially introduce hygiene issues. Furthermore, the physical / chemical measurement of alcohol level does not correlate directly with how drunk a person is turns of behavioural characteristics. For a given person, parameters such as tolerance levels, food levels, body size and so on may all impact the level of inebriation. Accordingly, people having a given measured level of alcohol may not all exhibit the same level of intoxication (drunkenness); rather, for such a given measured level of alcohol, some people may present as drunk, while other people may not present as drunk. Information about developments of automated alcohol dispensing systems in the hospitality industry can be found, inter alia, in patents such as the following: US11762966, US20160116181A1, US20230096235A1, US9915644, CN108961714A, and US20200118401A1. These developments have helped to bring about improvements in operational efficiency and customer personalization. Traditionally however, hospitality establishments still rely on human judgement for the assessment of intoxication, and this is prone to inconsistency and subjective biases. As a result, there may be concerns regarding safety, legal compliance, and operational efficiency with respect to alcohol dispensing systems, especially those in high-volume settings. The present application addresses the problem of enhancing automated drink vending machines to support the provision of alcoholic beverages having regard to the concerns set out above. Summary The invention is defined by the appended claims. A method and system for dispensing an alcoholic drink to a user are provided. The system may comprise at least one artificial neural network configured to receive, as input data, image data relating to: (i) the gait of the user; (ii) the pupil size of the user; and (iii) the Horizontal Gaze Nystagmus (HGN) of the user. The at least one artificial neural network is further configured to process the input data to produce an output which indicates whether or not the user is intoxicated, and if so, the system is configured to prevent dispensing of an alcoholic drink to the user. The method may comprise operating a drink dispensing system to handle a request by a user for an alcoholic drink from the drink dispensing system. The method comprises: receiving image data of the user; using the image data for measuring (i) the gait of the user; (ii) the pupil size of the user; and (iii) the Horizontal Gaze Nystagmus (HGN) of the user; processing by at least one artificial neural network (ANN) the measured (i) gait of the user; (ii) pupil size of the user; and (iii) Horizontal Gaze Nystagmus (HGN) of the user; generating an output from the at least one ANN to indicate whether the user is intoxicated; and instructing a drink dispensing device to deny or accept the request by the user according to whether or not the user is intoxicated. Brief Description of the Figures Various implementations of the claimed invention will now be described by way of example only with reference to the following drawings. Figure 1 is schematic diagram showing an example of a drink dispensing system as disclosed herein. Figure 2 is a schematic diagram showing an example of a user approaching the drink dispensing system of Figure 1. Figure 3 is a schematic diagram of a control module for use with the drink dispensing system of Figure 1. Figure 4 is a flowchart showing an example of a Reinforcement Learning (RLHF) process utilised by the system to refine or (re)adjust the assessment of intoxication level by an artificial neural network (ANN). Figure 5 is a flowchart showing an example of the Data Processing and Analysis process utilised by the system such as to refine and update the ANN. Figure 6 is a flowchart showing an example of the Environmental Adjustment and Reporting process utilised by the system such as to accommodate environmental factors when making decisions with the ANN. Figure 7 is a flowchart of the Decision-Making and User Interaction process utilised by the system to perform ongoing reinforcement learning and adaptation of the ANN. Figure 8 is a schematic diagram showing an example of a control centre operating in conjunction with multiple drink dispensing systems as disclosed herein. Figure 9 is a schematic diagram showing an example of a drink dispensing system having multiple artificial neural networks (ANNs) such as for determining whether or not to allow a drink to be dispensed to a given user. Figure 10 is a flowchart illustrating an example of operating or controlling an automatic drink dispensing system as described herein. Detailed Description The hospitality industry currently faces significant challenges in accurately assessing intoxication levels for alcohol dispensation. Traditional methods, such as breathalysers and manual sobriety tests rely on human judgement, leading to potential inconsistencies and subjective biases. There are systems that are accurate, such as a blood test, but these are invasive and not suited to the hospitality industry. This inconsistency in determining intoxication can result in the overservice of alcohol, exacerbating issues like antisocial behaviour and increasing the risk of accidents, legal repercussions, and safety concerns for establishments. Furthermore, these methods often prove to be intrusive and uncomfortable for customers, negatively impacting their experience. In high-volume settings like automated bars, the necessity for human intervention to verify a customer's sobriety status leads to operational inefficiencies, creating bottlenecks that slow down service and affect customer satisfaction. To address these challenges, an approach has been developed as described herein to introduce an innovative, automated system that utilises advanced technologies such as facial recognition, gait analysis, deep learning algorithms, and eye tracking, all at relatively low cost and based on non-invasive techniques (unlike other systems). This approach offers objective and consistent assessments of intoxication, significantly reducing human error and bias. The automation of intoxication assessments enhances operational efficiency, particularly in busy environments, by reducing or eliminating the number of manual checks. The non-invasive nature of the technology offers a more comfortable and privacy-conscious experience for customers. Moreover, by providing accurate and reliable intoxication assessments, the system aids establishments in adhering to legal standards and so helps to reduce any liabilities associated with over-serving alcohol. Ultimately, the approach described herein not only helps to improve the safety and compliance aspects of alcohol dispensation, but also significantly enhances the overall customer experience, fostering a safer and more enjoyable environment in hospitality settings. The approach described herein may be used, for example, to provide an automated, objective system for monitoring and controlling alcohol consumption, thereby extending the capabilities of existing systems in this context. As described herein, a self-service automated bar system may adeptly fuse machine learning, computer vision, and convolutional neural networks (a form of artificial neural network) to provide an accurate assessment of intoxication to regulate automated alcohol dispensation. Capitalizing on non-invasive, marker-less methods, an automated bar such as described herein may evaluate three primary intoxication indicators: gait (individual walking patterns), Horizontal Gaze Nystagmus (HGN), and pupil dilation. This bar system may integrate infrared and depth-sensing cameras, ambient sensors, and advanced algorithms from artificial intelligence and machine learning (AI / ML). These components work in combination with one another to provide a meticulous, multi-faceted evaluation of potential inebriation. This convergence of technology offers a holistic, non-intrusive and marker-free solution, using technological innovation to provide an ethical alcohol service in locations having an automated (autonomous) bar system (also referred to herein as a drink dispensing system). Accordingly, relying on breath samples and sobriety tasks as indicators of intoxication is constrained by a dependency on human judgment. This inherent subjectivity has introduced potential inconsistencies in determinations of intoxication, which may lead to the overservice of alcohol, potential escalation of antisocial behaviour, and the looming threat of legal and health complications. The existing development of self-service (automated) bars has led to an increase in operational efficiency and in user-centric engagement. Such machines (bars) may support a more tailored beverage dispensing experience, with certain systems being able to exploit user-specific data and inclinations for a tailored alcohol service. However, a fully autonomous system has been inhibited to date by the widespread use of human monitoring and intervention to ensure customers are in a fit state to consume alcohol. This monitoring introduces inefficiencies, particularly in fast-paced environments such as a festival, creating bottlenecks in which human assessors verify the suitability of each individual to access alcohol. In contrast, the system described herein supports determining a user’s intoxication level independently of human intervention with the aim of dispensing alcohol efficiently and responsibly. A bar system as described herein may advantageously combine a plurality of visual indicators such as gait analysis, HGN and pupil size assessments, to accurately determine the intoxication level of a target user. In an example, such a bar system may be configured to dispense a beverage to a target user depending on the comparison between (i) a determined level of intoxication (as determined by the AI / ML system) and (ii) a predetermined maximum threshold for the level of intoxication. The system may be configured to recognize a potential consumer's intent to purchase. The bar system may include one or more imaging devices (camera systems) which may be configured to monitor aspects such as a target user’s gestures, facial expressions, and proximity to the dispensing unit. The one or more cameras may have ergonomic design considerations to accommodate users of varied physical stature and ensure inclusivity in intoxication metric capture. The system may be configured to minimise or remove visual interference such as arising from overlapping actions, lighting fluctuations, and unexpected intrusions from unrelated subjects. For example, the system may be configured to reduce or remove visual interference from the captured images or data for a target user. The bar system may deploy a multi-dimensional strategy encompassing dynamic baseline establishment, spatial temporal filters, selective focus algorithms, adaptive thresholding, and a continuous feedback loop. The bar system helps to support compliance with regional and / or national legal standards pertaining to intoxication detection and data protection regulations, ensuring lawful operation across diverse jurisdictions. Accordingly, an automated alcohol dispensing system may be equipped with self-service features. Such a system synergizes artificial intelligence (Al) and machine learning (ML) to discern the intoxication status of patrons. The system's architecture may incorporate an in-depth evaluation of individual walking patterns (gait), Horizontal Gaze Nystagmus (HGN), and pupil dilation. Such multifaceted assessments help to foster an environment of responsible alcohol consumption, particularly in relation to autonomous drink dispensing machines (bars). Figure 1 is a diagram of an example of a drink dispensing system 10 (drink dispensing device) as disclosed herein. The dispensing system 10 includes a dispensing component 5 for providing a drink to the user (customer). The drink can be supplied in different formats according to the details of the machine. For example, in one format, the drink may be supplied in a closed container such as a can. In contrast, a second format may fill a drink into a suitable vessel (e.g. plastic cup) as part of the operation of the drink dispensing system. In some implementations, a given drink dispensing system may support both formats, with a particular format being selected according to one or more criteria, such as the type of beverage being dispensed, user preference, and so on. The dispensing system (machine) 10 further includes two user interfaces, a control interface 2 and a payment interface 4. The payment interface 4 supports payment for a drink (or drinks) that is / are being dispensed by the machine 10. The payment interface typically supports payment by credit card and / or by a suitable app running on a smart phone. The payment interface typically utilises a contactless (wireless) connection, for example based on near field communication (NFC), RFID payments from a wristband or similar, and / or Bluetooth, for communicating with a credit card or smartphone as appropriate. The payment interface may also be configured to accept cash payments. Typically, the dispensing system 10 may delay providing the requested drink(s) to a user until suitable payment has been made. The payment interface 4 may be provided with an advanced sensor suite to help secure transactional security. The control (user) interface 2 may be implemented using a touch screen or by any other suitable two-way user interface and allows a user to make a purchase from the drink vending machine 10. The control interface 2 therefore provides an interactive interface for completing user purchases and other engagements as required. For example, the control interface 2 may allow a user to search for drinks that are currently available from the machine and to select or specify a drink for purchase. The control interface 2 may also facilitate payment with the payment interface 4 as described above, which then allows the requested drink to be physically dispensed into the dispensing component 5. It will be appreciated that the aspects of the dispensing machine 10 described so far, in particular the dispensing component 5, the control interface 2 and the payment interface 4, are generally known to the skilled person for systems configured to dispense (typically non-alcoholic) drinks. As described above, there are additional concerns if a dispensing machine is used to provide alcoholic drinks. In some cases, the control interface 2 of the dispensing machine 10 may ask each customer to confirm they are 18 or over to purchase alcohol. The dispensing machine 10 of Figure 1 includes a camera 3 and this camera (or another imaging device, not shown in Figure 1) may be used to image both a customer and their associated ID, thereby allowing the age of the customer to be verified against the presented ID. However, while age verification can be incorporated as above into the operation of dispensing machine 10, the issue of denying an alcoholic drink to someone who is already drunk is generally more problematic. The dispensing machine 10 described herein may be used to address this issue. For example, such a dispensing machine may be provided with strategically placed imaging devices, such as a first imaging device 3 and a second imaging device 1. The imaging devices (cameras) 1 and 3 generally have fixed positions, although they may be movable in some implementations. In some implementations, the imaging device 3 shown in Figure 1 may be (at least partly) integrated into the control interface 2. It will be appreciated that according to the circumstances of any given implementation, the number of imaging devices may be one - e.g. where the functionality of imaging devices 1 and 3 has been combined into a single imaging device. Conversely, the number of imaging devices may be more than two, for example to acquire more comprehensive and reliable image data concerning potential users of the drink dispensing system 10. With reference to Figure 2, the imaging devices 1,3 capture visual indicators of a target user (subject) 6 who is approaching the dispensing device 10. In some implementations, the first and / or second imaging devices 1,3 may be configured to obtain infrared images as a supplement or replacement for obtaining the images. The first and / or second imaging devices 1,3 may also (or alternatively) be provided with one or more depth-sensors - this sensing is sometimes referred to as the ‘z-axis’ and is generally perpendicular to the image plane of cameras 1,3. The use of infrared imaging and / or depth-sensing may (for example) help the dispensing system 10 to operate more robustly, such as coping with varied lighting scenarios, for example, a low light level at nighttime. The use of infrared (or potentially thermal) imaging and / or depth-sensing may also help the dispensing system 10 achieve more precise three-dimensional skeletal data for a target user. The first imaging device 3 may, in some implementations, also perform additional functionality, such as age verification as discussed above. However, in other implementations, the first imaging device may be dedicated to preventing use of the dispensing machine 10 by someone who is drunk, in which case any additional functionality may be provided by a separate imaging device (not shown in Figure 1). The choice of having the imaging device 3 perform the additional functionality or having an extra imaging device for performing the additional functionality can allow for the particular circumstances, such as the degree to which the additional functionality is or is not compatible with the use of the first imaging device 3 for its primary purpose, namely to help prevent access to the dispensing system 10 by someone who is drunk. The dispensing system 10 may include a processing unit (not visible in Figure 1) which is responsible for processing an image acquired by the imaging device 3 to perform facial detection. Accordingly, the image device is positioned at an appropriate height to obtain a good image of a customer approaching or using the dispensing system 10. It will be appreciated that facial detection systems are known to the skilled person, see for example: https: / / en.wikipedia.org / wiki / Face_detection In the present context, the image is processed to identify a face of a subject and in particular to locate the positions of two eyes within the face. The processing system may determine whether the two respective eye locations of the subject are both at the same height - in other words whether or not a line from one eye to the other eye is horizontal. A human subject who is sober generally maintains their head posture so that their two eyes are horizontal, i.e. level with one another. In contrast, failure by a user to maintain their eyes in horizontal alignment may serve as an indication that the user might be drunk. The camera 3 may also be used to follow the gaze (view direction) of a user based on measuring the position of the pupil (including the direction in which the user is looking) as the user interacts with the control interface 2 of the dispensing system 10. For example, if a message for the user appears on the control interface 2, a person who is sober may notice and move (direct) the pupils to read the message more quickly than a patient who is less sober or drunk. The camera 3 may also be used to determine whether the pupils of a subject are dilated. Further information relating to blinking by the user may also (or alternatively) be captured by the camera 3. The measured blinking rate may be used, inter alia, to determine the pupil reflex time as the user interacts with the screen interface 2. Any (or all) of the above information, such as measurements of eye movement (tracking), eye orientation, eye blinking and / or dilation of the eyes of a user, may be utilised to help assess whether or not the user is intoxicated. When monitoring the pupils, the system may be measuring parameters such as gaze direction. The system may also measure the blinking rate of the user to work out pupil reflex time as the user works with the screen. The system may further assess how the user is interacting with the screen to identify potential signs of HGN or partial HGN. In general terms, HGN may be used to look for uncontrolled movement of the eyes, a condition which may be caused by intoxication (and / or by various other medical conditions). In practical situations, a user may exhibit partial HGN, whereby the user has some partial or limited control over the movement of their eyes. In the present context, the term HGN should be understood to incorporate / encompass such partial HGN (rather than necessarily involving complete loss of control of eye movement). In some implementations, the imaging device 3 may be used both to detect pupil size and to provide Horizontal Gaze Nystagmus (HGN) analysis. Further information about HGN can be found, inter alia, in “Horizontal gaze nystagmus: the science and the law: a resource guide for judges, prosecutors and law enforcement”, published 1 July 1999 by the US National Transportation Library, see https: / / rosap.ntl.bts.gov / view / dot / 1647. Regarding the estimation of pupil size, this may be implemented in an uncontrolled environment, in that the user head position may vary based on parameters such as pitch and / or yaw of the head. Such parameter variation makes it more complex to determine pupil size accurately in such various (and varying) conditions. Accordingly, when working with pupil dilation, gait, and (partial) HGN, the present system may take into consideration depth and ambient light to update the algorithms to adjust the Al model’s expectation and result. Such adjustment is based on a multitude of inputs from the user 6 as they interact with the system and may be conducted in a test format without determining specific marks (so that the adjustment does not impact the current order status). The first camera 3 may be implemented as a specialized facial imaging unit for obtaining images when a user’s face has a relatively high proximity to the dispensing system 10 (so that the processing of the image data from the first camera is not distracted, for example, by other faces in the background). The first camera 3 may also be provided with one or more ambient sensors, for example to measure lighting levels, the presence of rain, and so on. Such information may help to enhance performance for the evaluation of pupil detection and / or HGN evaluation, particularly under varying environmental conditions. For example, if light levels are low, then the pupils of a subject are likely to be dilated to increase light collection by the lens (without such dilation being an indicator of intoxication). The second camera 1 may be configured in use to capture the gait patterns of a user (subject). Figure 1 shows the second camera 1 attached to a mount above the top of the dispensing device 10. This additional height for the second camera 1 helps to provide a greater (higher) field of view of the gait pattern of a user 6 who is approaching the dispensing device (as shown in Figure 2). However, other implementations of the second camera may be able to capture a full-height image of the subject without providing such a high mounting for the camera. The gait analysis determines how the user is approaching the dispensing system - for example, walking, jumping or dancing. An Al model used for gait analysis may be sensitive to criteria such as the height and / or gender of the user to provide increased accuracy in determining whether or not a user should be able to acquire an alcoholic beverage. The gait analysis can contribute to the assessment of intoxication in a direct matter and also in an indirect manner. In the former, the analysis of gait is used to (help) assess intoxication based on the movement of the user. For example, the gait of a person staggering from side to side might be indicative of inebriation. In the latter, information about gait may feed into other instrumentation. For example, gait information may be utilised in conjunction with a gaze direction (such as discussed above). If the gaze direction as determined by the imaging device(s) 1,3 is aligned with a direction of movement of a user (such as may be measured as part of the gait analysis), then this may be indicative of greater control by the user, and hence there may be a lower likelihood of intoxication for such a user. Further information about gait analysis can be found, inter alia, in: "A Deep-Learning Approach for Identifying a Drunk Person Using Gait Recognition" by Park et al, published in January 2023, Applied Sciences, 13, 1390, available via: https: / / www.researchgate.net / publication / 367321944_A_Deep-Learning_Approach_for_ldentifying_a_Drunk_Person_Using_Gait_Recognition Note that the approach adopted in the above citation is based on gait energy imaging. Some examples of the implementation disclosed herein utilise landmarks on the body to recognise gait, since the use of such landmarks has been found to provide a more robust outcome in a high traffic environment. This is especially the case with the incorporation of visual imaging that provides a measurement of depth (distance) from a user. For example, the measurement of depth may be obtained by using cameras 1 and 3 (or any other suitable imaging device) to form a stereoscopic pairing. Such a configuration mirrors the binocular arrangement of human eyes, in that an assessment of depth or distance can be made based on viewing an object from two imaging systems that have a known offset from one another. Further information about gait analysis can also be found in “Age and sex differences in normative gait patterns” by Rowe et al, published in Gait &Posture, volume 88, July 2021, pages 109-115, available via: https: / / www.sciencedirect.com / science / article / abs / pii / S0966636221001934. This paper indicates that age and sex differences impact normative gait patterns - so that accounting for such factors supports the creation of a model with an enhanced ability to utilise gait in the assessment of whether or not a person is drunk. The assessment of gait may be further supported by a measurement of the height of a user, such as may be obtained (for example) by using the first camera 3 and / or the second camera 1. Accordingly, as part of the analysis of gait, the first camera 3 and / or the second camera 1 may be used to capture image data as a user (subject) 6 moves towards the dispensing system 10. Figure 2 schematically depicts camera 3 acquiring image data corresponding to the eyes of the subject 6. Such image data may be captured by the first camera 3 during the user approach to the dispensing system 10 and transmitted to the processing unit (not shown in Figure 2) for use in HGN evaluation. Figure 2 also depicts the second camera 1 acquiring image data corresponding to the full height of a user (subject) 6 who is approaching the dispensing system 10. This image data may then be transmitted to the processing unit for use in gait analysis. As mentioned above, the camera units 1 and 3 may be positioned strategically to help improve or supplement captured images, for example by obtaining depth information, pupil measurements, stereoscopic images and visual indicators, etc, as well as performing an analysis of a target user 6 via techniques such as gait analysis and HGN. In some implementations, the analysis of gait may take into account physical features such as height, gender and / or weight. These physical features can be determined (estimated) such as by using one or two imaging devices 1,3, and the estimated values for these physical features may then be fed into an ANN system that determines gait for a person having the (estimated) values for these physical features. Figure 3 illustrates an example of a control console 20 which is used by an operator of a drink dispensing system 10. The control console (panel) 20 is generally in a separate location compared with the drink dispensing system 10, but is in two-way communication with one or more such drink dispensing systems 10. This communication may be implemented by wireless and / or wired connections according to the circumstances of any given installation. In some cases, the control console 20 may be located and operated on the same site as the drink dispensing system 10; in other cases, the control console may be located and operated at a site which is remote from the drink dispensing system 10. In the latter case, there may be a central location of operators who provide support to multiple drink dispensing systems 10 installed at various different locations (venues). The control console 20 may be implemented as a portable device such as a touch pad, a tablet, a notepad, a portable (laptop) computer or the like. The control console 20 implements a control panel which may be used to receive information from a drink dispensing system 10 which is currently being monitored and controlled using the control console 20. The control console 20 may be further used to transmit control commands from the control console 20 back to the drink dispensing system(s) 10. The control panel of Figure 3 may include a video feed 12 from a drink dispensing system 10 which is currently being monitored by the control console 20. The video feed 12 may, for example, be obtained from the first imaging device 3 or from the second imaging device 1, or may be obtained from some other video source (not shown in Figure 1). The video feed 12 provides the person monitoring the drink dispensing system 10 with a visual view (and typically also an audio signal, see below) of the user 6 and his / her interaction with the drink dispensing system 10. The system may also generate a warning signal such as an amber beacon to alert operatives if someone is drunk; this warning signal might be displayed via the box (window) 11 of the control console 20. The drink dispensing system 10 and control console 20 may be configured to provide an audio channel (generally two-way) between the user 6 of the drink dispensing system 10 and the operator of the control console 20. In order to support such an audio channel, both the drink dispensing system 10 and the control console 20 may be provided with an audio input 18 facility (some form of microphone) and an audio output 19 facility (some form of speaker). In some cases, the first and / or second imaging devices of the drink dispensing system 10 may incorporate such audio functionality, e.g. microphone and speaker, to support the audio channel. In other cases, the drink dispensing system 10 may be provided with additional functionality (not shown in Figure 1) to implement the audio channel with the control console 20 comprising the microphone 18 and speaker 19. Similarly, the video link 12 between the control console 20 and the drink dispensing system 10 may act to provide (include) the audio connection for real-time communications between the operator and the customer. By way of example, a customer 6 may utilise such an audio and / or video real-time communications link between the control interface 2 and the system control console 20 to contact the operator. The customer may then use this link, for example, to inform the operator that there is a problem with the drink dispensing system 10, to ask how to perform certain functionality using the drink dispensing system 10, and so on. The video feed 12 may also be used to monitor usage of the machine, for example to detect if a customer might be trying to break into the drink dispensing system 10. The control console 20 further includes a window or box 11 provided as part of the user interface of the control console. In the present example, this box 11 supports various functionality, such as to allow an operator of the control console 20 to display information for or from a particular drink dispensing system 10 (in a typical installation, the control console 20 may support multiple such drink dispensing systems.). In some implementations of this functionality, the control console 20 may only communicate with a single drink dispensing system 10 at a time, hence box 11 may be used to select which drink dispensing system 10 the control console 20 is to communicate with. In other implementations, the control console may maintain ongoing communications with multiple drink dispensing systems 10. In this case, an operator may select to focus the control console 20 onto the control interface 2 (and the user 6 thereof). For example, box 11 of Figure 3 indicates that the control console 20 is currently showing information for, and interacting with, a machine #1 (which corresponds to a particular drink dispensing system 10 and a current user thereof). Box 11 further allows the operator of the control console 20 to handle conditions relating to usage of the drink dispensing system 10. A first such condition relates to age verification. As discussed above, age verification may in some cases be performed by using the first imaging device 3. In this case the box 11 may be used to image and / or record the outcome of the age verification (accept or refuse) as already determined autonomously by the drink dispensing system 10. In other implementations, the age verification may be determined or confirmed by the operator of the control console 20. For example, the operator of the control console 20 may use the information provided as part of the video feed 12 to determine whether age verification should or should not be confirmed, and to update the setting in box 11 accordingly. Box 11 is also shown in Figure 3 to include a “Hold” option. This option might, for example, be selected by the operator of the control console 20 (or set initially by the drink dispensing system 10). This “Hold” option might be activated if an ID document provided for performing age verification is somehow damaged. In such circumstances, the Hold option might indicate the need for additional human intervention to determine whether a user 6 has provided a valid (legitimate) ID document. Box 11 may also be used to handle the interaction between a drink dispensing system 10 and a user who is intoxicated and who therefore should not receive any further alcoholic drinks from the drink dispensing system 10. As described in more detail below, a processing unit located in the drink dispensing system 10 (or elsewhere) may use the imaging information acquired by the first and second cameras 1,3 to determine any such intoxication. This determination may be provided by an artificial neural network (ANN), or by any other suitable artificial intelligence (Al) or machine learning (ML) system. If the processing unit of the drink dispensing system 10 finds no clear indication that a user of the drink dispensing system 10 is drunk, then the user is accepted (as per the Accept option in box 11). This acceptance may be determined autonomously by the drink dispensing system 10 and then reported back to the control console 20, for example. In contrast, if the drink dispensing system 10 finds an indication that a user of the drink dispensing system 10 is drunk, then the user may be rejected (as per the Refuse option in box 11). The Hold option in box 11 may be selected instead of the Refuse option (for example) if this is a borderline case, and there should be human oversight of the final decision regarding whether or not the user should be able to receive an alcoholic drink from the drink dispensing system 10. In some implementations of the drink dispensing system 10, a user may be able to contest a decision by the Al system not to allow the user to purchase alcohol. In this case, the Hold option in window 11 might provide a UI facility for the user to contact a human operator with regard to the negative decision, such as through a voice and video connection provided by the drink dispensing system 10 (as discussed above). Accordingly, human operators, who may be housed at a distant locale, may retain the capability to endorse or veto commands or decisions in relation to intoxication which are being generated at (and applied to) a drink dispensing system 10 by the processing unit (such as based on the received image data and sensor data). This intervention by human operators is then fed back to the appropriate hardware (control console 20 and / or drink dispensing system 10) to implement the updated decision as to whether or not a given user is allowed to purchase an alcoholic beverage. This feedback also provides the overall system with reinforced learning protocols. In effect, the outcome provided by a human operator in respect of a given subject represents a label which is applied to the image data acquired by the processing unit. This labelled data can be used to enhance (train) the AI / ML system corresponding to the ANN (or similar) for incorporation into the processing unit. For example, the processing unit may indicate that a subject is drunk and hence should not be dispensed an alcoholic beverage, but after further investigation by a human operator (e.g. an operator of a control console 20), the human operator may accept that the subject is not drunk and hence can be properly dispensed an alcoholic beverage. In this case, the conflict between the initial decision (classification) of the ANN and the subsequent decision by the human operator may be fed back in order to update and refine the ANN. Conversely, a further investigation by a human operator may confirm the initial finding of the ANN in the processing unit. This agreement may again be used to provide labelled training data, which in this case will generally be used to reinforce the current configuration of the ANN (because the current configuration has produced a result which matches the result that a human would make in the same circumstances). Accordingly, the use of operator input provides a Reinforcement Learning from Human Feedback (RLHF) mechanism which can be deployed to enhance the accuracy of the ANN via machine learning (ML). This ML can be implemented in multiple different ways. For example, in some cases, live training may be performed, in which the labelled training data is fed directly and immediately into the ANN which is in operational use (in conjunction with the drink dispensing device 10) to update the model in the ANN. In another example, the labelled data may be fed to an off-line ANN in order to update the off-line ANN (where off-line indicates that the ANN is not currently being used in a production environment). At suitable intervals, the updated off-line ANN may be used to update or replace the live (production) ANN models which are used in conjunction with the drink dispensing devices 10. Irrespective of how the labelled training is fed back and applied to the ANN(s), the outcome is that the RLHF algorithm is adjusted and fine-tuned in response to the realtime operational data such as may be collected by the control console(s) 20. Note that in the above procedure, the system makes an autonomous determination regarding possible intoxication for most users. If the determination is positive, i.e. the user is not inebriated, the user can be dispensed an alcoholic drink without any involvement of a human operator. In contrast, a human operator is usually only involved if the determination is negative, i.e. the user is found by the ANN to be inebriated. This reduction in the number of human operators leads to a clear increase in efficiency and may provide better results by having a smaller pool of more experienced human operators. An example of a drink dispensing system described herein may incorporate an interactive interface 2 which may be configured in use to receive user instructions. This interface may be in communication with (or integrated with) at least one camera system. The camera system may be configured in use to detect pupil sizes and to provide Horizontal Gaze Nystagmus (HGN) analysis. A camera 1 which has higher (superior) positioning may be configured in use to capture user gait patterns. The captured data may be transmitted to a processing unit in the system for analysis of the user gait patterns. A specialized facial imaging unit 3 may be provided in closer proximity to users and may be combined with ambient sensors. This arrangement may help to provide good performance in pupil detection and HGN evaluations, particularly under varying environmental conditions. The system may comprise a payment mechanism 4 as described herein, which may be fortified with an advanced sensor suite, bolstering transactional security. The system may further comprise a dispensation component 5 configured in use to dispense a beverage to the user. The camera units 1 and 3 may be positioned strategically to provide for improved capturing of images and visual indicators of a potential user. These improved images may be used to help determine aspects such as such as gait analysis, HGN, and pupil measurements. The system may further comprise a control console 20) as described above, designed for remote feedback integration into the machine. This feedback may be integrated into the system's reinforced learning protocols. Human operatives, typically from a distant locale, may retain the capability to endorse or veto system commands, and may be closely involved in training the algorithmic model of the Al system. The control console may comprise a feedback loop 9, a primary control interface 8, and an ancillary apparatus 7. The camera(s), which typically have fixed positions, may be used to capture visual indicators of a target user. The camera system may comprise infrared and depth-sensing capabilities, whereby the system is able to operate under varied lighting scenarios and to support precise three-dimensional skeletal data acquisition of a target user. The system may be configured to capture data relating to visual indicators from cameras and (optionally) sensors of target users to provide a suite of artificial intelligence models. The system may be configured in use to utilise real-time object detection from the camera system, and to process the visual indicators detected, using the processing unit, to determine a level of intoxication of a user. This level may be compared to a predetermined maximum level of intoxication and such comparison may be utilised to determine whether or not the system is allowed to dispense a beverage to the user. Figure 4 relates to the operation of a drink dispensing system 10 such as described above. The system may be configured to capture data relating to visual indicators from cameras and optionally sensors of target users for use with a suite of artificial intelligence models. The system is configured to utilise real-time object detection from the camera system and to process the visual indicators detected using the processing unit to determine a level of intoxication of a user. This level may be compared to a predetermined maximum level of intoxication and this comparison may be utilised to determine whether or not the system should dispense a beverage to the user. More specifically, Figure 4 presents a flow chart illustrating an example of a process in which an ANN refines or readjusts its intoxication level assessments using a Reinforcement Learning from Human Feedback (RLHF) model. Background information about RLHF can be found, inter alia, at: https : / / en. Wikipedia. org / wiki / ReinforcementJearning_from_human_feedback. As shown in Figure 4, the process begins at operation 410 with the RLHF providing the AI / ANN system with rewards or penalties according to whether the AI / ANN outcome agrees with or differs from a human assessment. The RLHF may train the reward model directly from human feedback. For example, if a machine predicts the person is drunk, the machine will get a +1 or -1 (reward or penalty respectively) depending on whether the prediction was correct or not. This may involve the use of multiple models with a "ground truth model" to ensure predictions are correct if the level that the venue will accept is not crossed. At operation 420, the assessment algorithms are refined. For example, an algorithm may specify particular parameters (features) to implement a model, and the model is then used to make decisions (inference), such as to perform a classification. At operation 430, the AIANN system incorporates human guidance. For example, in the context of the present application, human guidance may be provided in the form of a human decision (classification) as to whether or not a subject is intoxicated. This can be considered as a form of training input for the AI / ANN system. At operation 440, the parameters of the AI / ANN model (system) may be evaluated and adjusted, for example based on the human guidance. In other words, the values of the model parameters are refined so that the model outcome (in terms of classification) is closer to or coincident with the outcome that would be obtained from a human decision-maker. At operation 450, the AIANN system incorporates continuous learning and adaptation. Note that such learning may be off-line or on-line (as discussed above). Figure 5 further relates to the operation of a drink dispensing system 10 such as described above. Figure 5 presents a flow chart which illustrates an example of how data captured by the camera system (and then anonymized data) is processed by the processing unit. The processing unit may apply processes such as convolutional (artificial) neural network analysis, sequential data processing with Long short-term memory (LSTM) networks and / or generation of recommendations based on determined intoxication levels. Thus the processing of Figure 5 commences with operation 510, in which data (such as images, sensed data, and / or financial information provided as part of a sales transaction) is anonymized and encrypted into a file. At operation 520, the data is pre-processed and a feature extraction is performed. These features represent particular aspects which may be determined from an observation of an entity under investigation, and which provide a succinct and effective representation of the properties and state of the entity. At operation 530, the extracted features are analysed using an AI / ML system comprising one or more convolutional neural network (CNN). A CNN is a particular form of (artificial) neural network and is often used for handling images and similar data. A CNN has multiple layers of network nodes, each layer having an associated depth, and deeper layers in the neural network generally correspond to larger-scale features in the image (or other data set). At operation 540, sequential data processing is performed using LSTM, which is a tool for enhancing certain aspects of the performance of a neural network (especially in respect of small but persistent features). Further information about an LSTM tool can be found at: https: / / en.wikipedia.org / wiki / Long_short-term_memory. At operation 550, recommendations are generated and validated. In the present context, a recommendation would relate to whether or not to serve an alcoholic drink to a potential customer, based on the level of intoxication (if any) as detected by the ANN. Such a recommendation is obtained by applying received data (features extracted from images, etc) to the ANN(s) of the drink dispensing system 10. At operation 560, the model (such as implemented by the ANN) is updated with any new insights. This represents a form of training for the model, whereby the output from the model at operation 550 is validated (trained), typically by comparing the model output against a human analysis and output based on the same input data. If there is a discrepancy between the output from (i) a human operator, and (ii) the model running on the ANN, the model can be further trained (adjusted) using a form of back propagation to provide better alignment between the output from the human and output from the model running on the ANN. Figure 6 further relates to the operation of a drink dispensing system 10 such as described above. As shown in Figure 6, the system may be configured in use to apply an environmental adjustment and reporting process. Such a process may involve generating detailed reports on the intoxication levels of one or more user(s) in conjunction with environmental factors. The bar system may be configured in use to adjust its intoxication levels, as determined by the processing unit based on such environmental factors. The system may comprise one or more sensors (for example ambient sensors) for recording environmental factors. In some cases, the environmental factors may also (or alternatively) be acquired using the first and / or second cameras 1,3 (with or without additional sensors). The processing unit is configured in use to receive recorded environmental factors from the sensor(s) and / or imaging systems 1,3 to utilise these factors to adjust intoxication levels determined by the camera(s) 1,3. The system may be configured in use to generate detailed reports on the intoxication levels of one or more user(s) having regard to such environment factors. By way of example, one or more environmental factors may relate to the level of wind and / or the level of darkness. If there is a strong wind, a user may be less steady on his / her feet, even if they are sober and have consumed little or no alcohol. A similar circumstance may apply to darkness, when a user who is sober might still move more cautiously over the ground, and possibly trip over an unseen item. Accordingly, measurements relating to environmental factors such as wind and darkness may be incorporated into the assessment by the ANN model of sobriety or inebriation. The particular processing in Figure 6 shows an example of using one or more environmental factors to help in the assessment of sobriety or inebriation. In operation 610, the ambient sensors (or any other suitable device(s) for determining environmental factors) are activated. In operation 620, these sensors are utilised to help assess environmental factors which may impact the assessment of intoxication (if any). At operation 630, the system generates a report which reflects the above circumstances, for example, the state of the sensed environment factor(s) and also other indicators that may be used to assess the level of toxication. At operation 640, an unsupervised learning process is performed to identify emerging patterns (where unsupervised learning is performed without human indication of a correct output for each instance). For example, in the above scenario in which the system measures environmental factors such as wind and / or darkness, the system may discern that such features may be associated with an increase in certain behaviours (such as walking more carefully), whereas other indicators might indicate that the user is not intoxicated (such as rapid operation of the drink dispensing system without any indication of confusion). This can then lead to learnt behaviour (algorithms) within the ANN model regarding the impact of such environmental factors on the assessment of intoxication. At operation 650, the algorithms which are used to adjust the impact of environment factors on the assessment of intoxication can be refined, based for example on the information collected at operation 630 and then analysed at operation 640, to provide a more reliable and robust assessment of sobriety. Figure 7 further relates to the operation of a drink dispensing system 10 such as described above. The system may be configured to allow a third party, i.e. not the system 10 or the user 6, to submit additional input relating to outcomes determined by the system (for example to provide human approval). The system may, for example, include a control unit for use by a third party. Figure 7 illustrates such a process, in which the system is configured in use to make a decision based on a determined level of intoxication, and in which the system is further configured to receive input from a human in response to the decision of the system. The system is configured to adjust, where necessary, the outcome and / or decision determined by the processing unit on receipt of an input provided by a human (for example, via a control console 20), leading to a final outcome. In the example flowchart of Figure 7, at operation 710 a user (customer) 6 approaches a drink dispensing system 10, as shown also in Figure 2, In operation 720, the system identifies the user, for example via a driving license, phone app, RFID tag, etc, and initiates an interaction between the dispensing system 10 and the user 6. At operation 730, relevant data concerning this interaction (transaction) between the dispensing system 10 and user 6 is captured and anonymized (consistent with data protection laws, etc); this broadly corresponds to the analysis of operation 510 in Figure 5. At operation 740, in-depth data processing and analysis are performed. In general terms, this analysis collects the data obtained relating to this interaction, and feeds the data as an input to the ANN model(s), which then determine a corresponding output. At operation 750, intelligent decision-making is performed and the result of this determination is passed to the user 6. For example, the ANN model after the analysis of operation 740 determines whether or not the user shows signs of intoxication, and the decision-making at operation 750 then proceeds on the basis of this determination. In particular, the decision-making determines whether or not the drink dispensing system 10 should provide an alcoholic drink to a according to the perceived level of intoxication of the user, and the user is then informed accordingly, for example by the control panel 2 (see Figure 1). At operation 760, results from the drink dispensing system 10 are integrated and compiled into a data set which can then be assessed by the drink dispensing system 10 and / or the control console 20 (and / or any other suitable component or system). At operation 770, environmental factors or data may be adaptively adjusted based for example on the compiled data set generated at operation 760. At operation 780, the data and operation of the drink dispensing system 10 is further monitored on an ongoing basis to provide continuous reinforcement learning and adaptation, which may then be used, for example, to update the ANN model according to experience of how the ANN model performs in an operational environment (field). Accordingly, by way of example, a system described herein may be configured such that the algorithm adjusts and fine tunes in response to real-time object detection and detailed skeletal mapping. As discussed above, Figure 4 illustrates a process in which the system is configured in use to refine or readjust its Intoxication level assessments using a RLHF model. The system is configured to use Al to assimilate anonymous rewards, comparisons and / or human guidance. In Figure 5, the data captured by the camera system (for example anonymized data) may be processed by the processing unit. The processing unit may apply processes such as convolutional neural network analysis, sequential data processing with LSTM networks and / or generation of recommendations based on determined intoxication levels. In Figure 6, the system may be configured in use to apply an environmental adjustment and reporting process. The system may be configured in use to adjust its intoxication levels, as determined by the processing unit, based for example on environmental factors. The system may comprise one or more sensors (for example ambient sensors) for recording environmental factors. The processing unit is configured in use to receive recorded environmental factors from the sensor(s) and to utilise these to adjust intoxication levels determined by the one or more cameras. The system may be configured in use to generate detailed reports on the intoxication levels of one or more user(s), optionally in addition to environmental factors. The system may be configured to allow for third parties to input additional responses (for example to receive human approval) in relation to outcomes determined by the system. The system may for example include a control unit for use by a third party. Figure 7 shows an example of a process in which the system may be configured in use to make decisions based on determined levels of intoxication, and in which the system is further configured to receive human input in response to the decisions of the system. The system may be configured to adjust, as appropriate, the outcomes and / or decisions determined by the processing unit upon receipt of human inputs (via for example a control unit / control console) leading to a final outcome. A system described herein may advantageously combine a plurality of visual indicators, such as gait analysis, HGN and pupil size assessments, to determine accurately the intoxication level of a target user. Such a system may be configured to dispense a beverage to a target user depending on a comparison between (i) a level of intoxication estimated by the ANN(s) and (ii) a preset maximum threshold for the level of intoxication. The visual indicators of a user 6 may be captured by one or more cameras 1,3. Such cameras may be designed or configured to accommodate users of varied physical stature, thereby helping to provide inclusivity in the determination of any intoxication. The system may be configured to recognize a potential consumer's intent to purchase. The camera(s) 1,3 may be configured to monitor the potential consumer's (user’s) gestures, facial expressions, and proximity to the dispensing unit. The system may be configured to minimise or remove visual interferences (such as overlapping actions, lighting fluctuations, and unexpected intrusions from unrelated subjects) from the captured images for the potential user. The system may for example employing a multi-dimensional-strategy encompassing one or more of dynamic baseline establishment, spatial temporal filters, selective focus algorithms, adaptive thresholding, and a continuous feedback loop. The system may be configured to comply with regional legal standards pertaining to intoxication detection and / or data protection regulations, thereby supporting lawful operation in diverse jurisdictions. Figure 8 is a schematic diagram showing an example of a control centre 755 operating in conjunction with multiple drink dispensing systems 10 as disclosed herein. It will be appreciated that the configuration and architecture shown in Figure 8 are provided by way of example, and many alternatives could be adopted according to the circumstances of any given installation. The control centre 755 is shown in Figure 8 as including control consoles 20A, 20B, 20C, (referred to collectively as 20). Each of the control consoles may be implemented, for example, in the same general manner as control console 20 (see Figure 3). Although Figure 8 depicts the control centre 755 as having multiple (3) control consoles 20, in some installations there might be (i) only one or two control consoles 20, or (ii) (possibly substantially) more than three control consoles 20. In addition, Figure 8 shows a single control centre 755, but some implementations may utilise multiple control centres 755, for example, different control centres might potentially be allocated to different geographical regions (countries, states, etc). The control centre(s) 755 may be used, for example, to perform operational control (a first task) and to provide computer implementation (a second task). If there is more than control centre 755, then tasks may be split between different control centres, or some (at least) of the control centres 755 may be used to perform both tasks. Figure 8 further shows two vending locations 758A, 758B (but again it will be appreciated that there may be a single vending location (or more than two such vending locations). The vending location(s) 758A, 758B are connected to the control centre(s) 755 via cloud communication 701, which may utilise any suitable network(s), such as the Internet, mobile telephone networks, and so on. Each vending location 758A, 758B may include one or more drink dispensing machines 10, such as depicted and described above with reference to Figures 1 and 2. As shown in Figure 8, the number of drink dispensing machines 10 may vary from one vending location 758A to another vending location 758B. For the first task, operational control, the control centre 755 (more particularly the control consoles 20 within the control centre) may provide human monitoring of the vending locations 758A, 758B and may potentially make an intervention according to the status of each dispensing device 10. For example, a user 6 at a given dispensing machine may initiate an audio (or telephone) communication with the control centre 755 if a dispensing device does not appear to be working properly (or an error message is displayed). A human monitor at the control centre755 may have the expertise and authority to remotely access the dispensing device 10 over cloud network 701 to try and rectify the error (or, if appropriate, close down the dispensing device pending action by a maintenance engineer). As another example of operational control, a dispensing system 10 may have concluded that a user 6 is inebriated and so not entitled to purchase an alcoholic drink. The user 6 might ask a human monitor at a control console 20 (e.g. 20B) to review this refusal if they consider the finding of the dispensing system to be incorrect, or if there are special circumstances that the user thinks have not been taken into consideration. For the second task, some or all of the code utilised for the computing implementation, such as to support the ANN systems, may be run at the control centre 755, or perhaps elsewhere in or accessible via the cloud 701. More particularly, the control centre 755 may provide some or (nearly) all of the computing resources to implement the drink dispensing described herein. For example, standard functionality of a vending machinelO might be implemented locally at the drink dispensing system, but additional functionality as described herein, such as in relation to the neural networks, etc, may be run on the control consoles 20A, 20B, 20C or on other computer systems located at or accessible to the control centre 755. Figure 9 is a schematic of a system 800 disclosed herein for automatically dispensing alcoholic beverages. It will be appreciated that the configuration of Figure 9 is provided by way of illustration, rather than limitation, and may variations may be adopted according to the circumstances of any given implementation. In broad terms, the system 800 is arranged to refused to supply an alcoholic drink to a user who presents as intoxicated and to supply an alcoholic drink to a user who does not present as intoxicated. Figure 9 shows a system 800 including one or more processing units 814 which include control software 848. In practice, a processing unit 814 may comprise multiple hardware and software components for performing different computational tasks. For example, the control software 848 may comprise a set of one or more program components which may operate together or separately from one another. The processing unit 814 may (for example) be incorporated into a drink dispensing system, such as the drink dispensing system 10 of Figures 1 and 2. The processing unit 814 may be implemented by any suitable computing device such as a tablet, notebook, laptop, etc, where such devices may be adapted for inclusion in the drink dispensing system 10. The processing unit 814 may include one or more graphics processing units (GPUs) to support processing performed by or in relation to the ANNs 818a, 818b, etc (denoted generically as 818) which are implemented within the processing unit 814. The control software 848 may run on the computing device to provide the desired functionality described herein. The processing unit 814 is further shown in Figure 9 to include three artificial neural networks (by way of example). Although Figure 9 illustrates the use of artificial neural networks 818, other implementations may use any suitable form of artificial intelligence (Al) and / or machine learning (ML) for performing the functionality described herein. Furthermore, ANNs should be understood broadly as any computing system based on a neural network configuration. The processing unit 814 is connected to a set of sensing devices 860, for example via a local communications facility 831. Such a connection may be implemented by any suitable communications link, such as Bluetooth, wired (e.g. USB), WiFi, and so on. Note that different sensing devices 860 may utilise different forms of communication link for connecting to the processing unit 814. In some implementations, the sensing devices 860 and the processing unit 814 may be located in the same device or system, such as a drink dispensing system 10, and hence may be connected to one another by internal wiring of the drink dispensing system 10. The sensing devices 860 are shown in Figure 9 as including one or more camera(s) 867, one or more sensors 868 and an infrared illuminator 869. The sensing devices are generally installed into the drink dispensing system 10 or are closely associated with the dispensing system 10 to obtain high quality sensing information relating to a user 6 and / or surrounding environment. The set of sensing devices 860 included in system 800 may vary according to the circumstances of any given implementation. In the example of Figure 9, the sensing devices 860 may comprise at least two imaging devices such as cameras 1 and 3 shown in Figures 1 and 2; these images may be configured to provide still and / or video imaging. A first camera 3 may be located in the drink dispensing system 10 directly adjacent to the control interface 2 to capture images of the face of a user 6 who is currently operating the dispensing system 10. The imaging from camera 3 is suited, for example, for determining pupil size and HGN. A second camera 1 may be located in a raised position above the main drink dispensing system 10 and is suited, for example, for determining gait. In some implementations, the camera 1 and / or the camera 3 have near infrared sensitivity, for example, sensitivity across some or all of the wavelength range 800-1200 nm. In other implementations, a separate infrared camera may be utilised in addition to cameras 1,3. The sensor(s) 868 may be used to measure environmental factors such temperature, wind speed, general light levels, noise level, and so on. These environmental factors may be provided as input to the ANNs 818 to support a more refined analysis and hence a better estimate of intoxication. For example, the general lighting level may impact pupil size, which is one of the features used for the ANN modelling. In this context, the variation in background light level can be regarded as noise. The processing unit may use information about the background light level to help to estimate to what extent a measured pupil size is attributable to background light level (noise) and to what extent the measured pupil size is attributable to user intoxication (in effect, the signal). This availability of the background light level therefore allows a better estimate of the level of user intoxication. The sensing devices 860 may further include an infrared light source 869 in or associated with the drink dispensing device 10. This infrared light source is generally not visible to humans, but provides illumination to help a camera having infrared sensitivity obtain better images, e.g. of the pupils. (Another possibility is to provide illumination in the visible range, although this is more likely to alter the pupil size, similar to the noise discussed above). The drink dispensing system 10 generally includes one or more input / output facilities 882 for a user. The user input / output facilities may comprise or include a control interface 2 such as discussed above in the example of Figures 1 and 2. In Figure 9, the user input / output facilities 882 may be implemented by a two-way audiovisual link together with microphone and speaker 886. In some implementations, the audiovisual link may include a camera 867 and a display screen to support video communications. In Figure 9, the display screen may comprise a touchscreen 887 which allows the user 6 to enter information, commands, and so on for transmission to the processing unit(s) 814. The user input / output facility 882 may also be implemented by a telephone (smart phone) app 888 based on standard functionality of a smart phone to transmit and receive real-time audiovisual communications, control data, and so on. Note that the user input / output facility 882 may be connected to the processing units 814 by a local connection 831 as discussed above, and / or by a remote connection 832 (such as may be utilised by Phone App 888 interacting with a remote website). The system 800 further comprises a remote console 820 and a remote computing platform having one or more servers 840. Both the console and the server(s) may be connected to a processing unit 814 within a drink dispensing system 10 by a remote communications network (such as the Internet) 832. In some implementations, the server(s) 840 and the remote console 820 may be integrated into a single system at a given location. In other implementations, the remote console 820 is separate from server(s) 840. For example, the remote console 820 may correspond to the control console 20 shown in Figure 3 and be used to host human support operators who monitor and interact with users of various drink dispensing systems 10 which may be installed in multiple different locations. The server 840 and associated software 847 may correspond to the control centre 755 of Figure 8. Similarly, the remote communications facility 832 may correspond to or be implemented by the cloud computing platform 701 of Figure 8. Furthermore, the system 800 may include vending locations 758 hosting one or more drink dispensing system 10 such as illustrated in Figure 8, and a drink dispensing system 10 may, for example, include at least one processor unit 814, sensing devices 860 and user input / output 882. The server 840 shown in Figure 9 has ANNs 818d, 818e (by way of example). In practice, the ANNs 818 may be located on processing units 814 and / or server(s) 840 according to the circumstances of any given implementation (without limitation to the particular allocation of ANNs shown in Figure 9). The server(s) 840 further include control software 847, which is somewhat analogous to the control software 848 in the processing unit(s) 814 and may comprise multiple components for performing a range of different tasks. For example, the control software (848 and / or 847) may be used to oversee the operation of one or more ANNs 818 (located at the server 840 and / or the processing units 814) to provide a final estimate of intoxication for a user. Similarly, the control software (848 and / or 847) may be used to further train the ANNs as additional training data is acquired and applied to the ANNs. In some implementations, the system 800 may comprise a single ANN that takes in all available data, such as imaging, sensor signals, and so on, and applies this data to the single ANN to obtain a single output estimate for intoxication. In contrast, the implementation of system 800 shown in Figure 9 comprises multiple ANNs, wherein each ANN models a particular component of the overall model. In particular, the system 800 of Figure 9 comprises at least one ANN for determining gait, at least one ANN for determining pupil size, and at least one ANN for determining HGN. This use of different ANNs focussed on different physical aspects of a user is considered to provide better estimates of intoxication. Each ANN is configured to receive appropriate input data - for example, an ANN relating to pupil size receives and utilises images of the eyes of a user, whereas an ANN relating to gait receives and utilises images of the whole body of the user. Each ANN therefore receives the relevant input data and applies this input data to the weighted model of the ANN to determine an output for that ANN. This output can then be combined with outputs from other ANNs to obtain an overall result. For example, some implementations may provide a simple binary output from each ANN: “yes” - intoxicated, “no” - not intoxicated. A threshold may then be set to determine a final outcome from the overall set of ANNs. For example, the threshold may hold that a user is intoxicated only if each of the ANNs give a “yes” output, or the threshold may hold that a user is intoxicated if any one of the ANNs gives a “yes” output. In other implementations, the output from each ANN may represent a score from a range, say between 0 and 20 (although different ANNs may be defined to have different scoring ranges). A higher score would indicate a greater likelihood of intoxication. By way of a simplified example, there may be three ANNs (one for gait, one for pupil size, and one for HGN) to give a maximum total score of 60. In this case the threshold could be set to an appropriate value, say 40, and if the three scores total 40 or more, the user is considered to be intoxicated, whereas if the three scores total less than 40, the user is considered not to be intoxicated. It will be appreciated that the above approach for making a linear combination of the scores from individual ANNs may be replaced by other more complex approaches - such as the use of nonlinear combinations (which could be implemented using an additional ANN). In some cases, an approximate error range of the individual scores and / or the final output may be determined. Such an error range may be estimated (for example) by repeating the analysis multiple times, but using ANNs with slightly different weights and / or with slightly different input data (measurements). These multiple different weights and / or different inputs can be generated by applying noise to the initial weights or inputs. If such an error range is available, this may also be taken into account when assessing the level of intoxication. For example, returning to the above example with a maximum score of 60 and a threshold value of 40, then a finding of 43 with an error range of ±5 might be regarded as not clearly intoxicated, whereas a finding of 43 with an error range of ±1 might be regarded as intoxicated. In some implementations, the different ANNs may be set to have different weightings according to the particular circumstances. For example, one property, such as gait, might generally be the most sensitive to intoxication. In this case, there might be a scoring range of 0-15 for each of pupil size and 0-15 for HGN, and a scoring range of 0-30 for gait. The maximum total level therefore remains 60, but different ANNs are weighted to make different contributions to this model. The weightings applied to the different ANNs may also vary according to the measured values relating to environmental parameters. For example, if a sensor detects rainfall, the system 800 might reduce the weighting applied to gait, since users are more likely to slip in the wet conditions (without being intoxicated). As another example, if the light level if rapidly changing (such as during a musical performance with lighting), the system 800 might reduce the weighting applied to pupil size, since the pupil size might vary because of the show lighting (rather than due to intoxication). Figure 10 is a flowchart illustrating an example of operating or controlling an automatic drink dispensing system 800 as described herein. The processing of the flowchart commences at operation 1010 with receiving image data of a user. In general, a user 6 may have requested an alcoholic drink from the system 800 (which may incorporate one or more drink dispensing systems 10 such as shown in Figures 1 and 2). In order to determine whether to allow or reject a request for an alcoholic drink, the image data is acquired using one or more camera(s) 867 (and potentially one or more other sensors 868) to support an estimate concerning the level of intoxication (if any) of the user. The acquired image data (or data derived from the image data) is then provided to (received by) a suitable device, such as a processing unit 814. At operation 1020, the received image data is used to measure current properties / parameters of the user, at least gait, pupil size and HGN. These measured properties / parameters are then utilised to assess the level of intoxication of a user. In general, this typically involves a two-stage process: firstly, the image data assesses (measures) the gait, pupil size and HGN, and then secondly, the measured gait, pupil size and HGN are used to determine the intoxication level of a user. Note that in some configurations, the above two-stage process may be implemented as a single stage. In operation 1030, artificial neural networks (ANNs) 818 are utilised for the first and / or second stage. Typically, the ANNs are used for the second stage, namely to derive an estimate of intoxication from the measurements of gait, pupil size and HGN. Accordingly, the measurements for gait, pupil size and HGN may be regarded as features extracted from the image data which are then provided as inputs to the ANNs. The trained structure of the ANN(s) processes (analyses) the inputs, in effect applying measured data sets (pupil size, gaze, HGN) to the ANN model. In operation 1040, the analysis by the ANNs is used to generate an output regarding the state of intoxication of the user. By way of example, the model output may be expressed as a classification having two possibilities: (a) the estimated intoxication according to the ANNs is below a threshold, or (b) the estimated intoxication according to the ANNs is above the threshold. At operation 1050, the output generated by the ANN(s) is used to provide a control instruction to the drink dispensing system 10. The instruction may permit the drink dispensing system to dispense an alcoholic drink to the user (in the case of option (b)) or may instruct the drink dispensing system 10 not to dispense an alcoholic drink to the user (in the case of option (a)). It will be appreciated that the procedure of Figure 10 utilises ANN(s) that are production systems, i.e. they have already been trained. As discussed above, see for example Figure 5, these production ANN systems may be generated by using training data to set the internal weights of the neural network. During such training, a drink dispensing device 10 is used to measure user parameters such as gait, pupil size and HGN, but the yes / no determination on intoxication is made by a human monitor or supervisor (who is generally assumed to make a correct assessment). During the training phase, the ANN model is therefore adjusted to align the model output with the human decisions. Note that further training data may become available after deployment of system 800. For example, as discussed in relation to Figure 3, a human support operator located at remote console 20 (corresponding to remote console 820 in Figure 9) may override the finding based on the ANNs that the user 6 is too intoxicated to be permitted another drink. The system can use information about such an override to update the ANNs 818 to align better with the human finding. In some cases, this updating (adjusting) of the ANNs may be performed on an ongoing basis to the production ANNs. In other cases, the updating may be performed initially on a development system for subsequent deployment onto the live (production) ANNs. While Figures 9 and 10 and associated description provide various examples of a system 800 for automatically dispensing alcoholic beverages, many other implementations may be utilised according to the circumstances of any given Installation. For example, with reference to Figure 10, the processing shown in Figure 10 may be supplemented by various additional processes as described herein and / or the classification may be confirmed by a human supervisor, which also allows machine learning to be perform to improve the accuracy of the classification (such as described above). Furthermore, the operation 1010 may be performed locally in the presence of a user 6, such as by a drink dispensing system 10, or by a device which is installed within or in close proximity to drink dispensing system. This allows the camera(s) 867 to image the user 6 and obtain information from any sensor(s) used to support the operation of a dispensing system 10 by providing information on ambient conditions. This acquired data is then received for example by processing unit 814 and / or by a server 840 to help assess the level of intoxication for a user. The remaining operations 1020-1050 in Figure 10 are not restricted in performance to a location on or close to a drink dispensing system. Rather, each operation 1020-1050 may be performed by one or processing units 814 installed locally to the dispensing system 10 and / or by one or more remote processing facilities, such as server 840. The final operation 1050 may then involve sending an instruction, such as from a local processing unit 814 or from a remote server 840, to a control facility of the dispensing device 10 which then implements the received instruction to dispense or not dispense an alcoholic drink to the user 6. Accordingly, it will be appreciated that there are many implementations for performing the operations set out in Figure 10, and a particular implementation may be selected because of current circumstances of the overall system 800 shown in Figure 9. An automated intoxication detection system 800 as described herein may be used for automated bar environments. Such a system may comprise a processing unit 814 configured to analyze and assess visual and physiological metrics related to intoxication. The processing unit may further comprise a machine learning algorithm 818 trained to correlate anonymized user metrics with levels of intoxication. The processing unit is generally configured to receive images and / or data, such as one or more visual markers, relating to one or more visual and / or physiological metrics associated with intoxication of a target user 6. The system may further comprise a camera system comprising at least one camera 1,3 located at one or more predetermined locations. The camera system may be equipped with infrared and / or depth-sensing capabilities and may be configured in use to detect one or more visual markers indicative of intoxication of one or more users. Such visual markers may comprise a combination of: Gait Analysis, Horizontal Gaze Nystagmus (HGN) and pupil dilation. The camera system may be in communication with the processing unit 14 to provide visual and physiological metrics determined from the detection of the one or more visual markers. The camera system may be in communication with or integrated within an interactive user interface 2 configured in use to provide focused visual data capture from target users interacting with the system. The camera system may include one or more sensors 868 configured in use to determine, detect and / or measure one or more of the visual markers (characteristics) in addition or as an alternative to a camera. The one or more sensors may be in communication with the processing unit to provide visual and / or physiological metrics determined from the detection of the one or more visual markers. The camera system and / or the processing unit may be configured in use to employ noise reduction algorithms to filter out (or reduce) at least one background activity. Such an approach is adopted in order to isolate one or more target user's visual and / or physiological metrics, for example from captured visual and / or physiological metrics. The noise reduction algorithm may be configured to filter out or remove (or reduce) at least one background activity in, for example, crowded settings. (Such background activities to be filtered out may comprise persons and / or movement occurring within the field of view of the camera system which are not related to the target one or more users). The processing unit may be configured in use to leverage convolutional neural networks (CNNs) for visual data interpretation and long short-term memory (LSTM) networks to analyze sequential patterns in the one or more visual indicators. In some cases, the system may further comprise a Reinforcement Learning from Human Feedback (RLHF) mechanism configured in use to allow the system to refine or readjust detection capabilities based on real-world feedback and data. The processing unit may be configured in use for collaborative operation with one or more external systems. For example, an external system may be a closed-circuit television (CCTV). The bar system may therefore be configured to allow for retrospective analysis of the interaction with a target user. This may for example help in determining if a target user subsequently operated a vehicle post-interaction with the system. The system may further comprise at least one server 840 and at least one encrypted communication pathway configured in use to enhance data privacy and security during an intoxication assessment. The bar system may further comprise a secondary user interface (console) 20 configured to allow (inter alia): (i) human supervisors to oversee system data in real-time; (ii) human supervisors to contribute to RLHF inputs; and / or (iii) human users to provide details of on-site or offsite alcohol service decisions. Accordingly, a Self-Service Automated Bar System 800 is disclosed herein which may employ advanced machine learning and artificial intelligence to perform a meticulous and anonymous analysis of intoxication levels of one or more users to provide for a responsible and precise alcohol dispensing service. A system such as described herein may comprise a plurality of cameras 867 and (optionally) one or more sensors 868 in communication with a processing unit 814. An artificial intelligence (Al) system 818 may then be used to communicate and / or integrate with a point-of-sale system 10 and / or a wearable device to provide a comprehensive and sophisticated approach to responsible alcohol dispensing in automated environments. Such a system is configured to integrate detected HGN data and pupil dilation with gait analysis of a user as determined by a camera system. This combination of one or more such visual markers results in a system with heightened accuracy and reliability in determining the intoxication level of a user. The identification, measurement and / or determination of Horizontal Gaze Nystagmus (HGN) by the camera system (and / or sensor(s)) and the pupil size analysis enable the system to accurately discern the intoxication levels of a target user such as by detecting involuntary movements of the eye and variations in pupil size which are (typically) indicative of alcohol impairment. The system may further comprise a payment mechanism 4 for actuation by a user to select and pay for a selected beverage. The payment mechanism is configured to communicate with the processing unit. The system may further comprise a dispensing mechanism 5 which is configured to communicate with the payment mechanism and the processing unit. The processing unit may be operable such that, on receipt of payment for a beverage and on receipt of data confirming that the measured intoxication level of a target user is below a predetermined maximum threshold, the processing unit actuates the dispensing mechanism to dispense the beverage. The processing unit may also be operable, on receipt of payment for a beverage and on confirmation that the intoxication level of a target user is equal to or greater than a predetermined maximum threshold, to prevent an alcoholic drink from being dispensed by the dispensing mechanism. Accordingly, the present approach integrates a range of advanced technologies to create a comprehensive solution. Major features include depth analysis cameras and computer vision for face tracking, determining the optimal movement for performing a gait analysis by measuring a user’s distance from a screen 2. Advanced deep learning algorithms are used for individual tracking, assigning unique IDs to patrons and enabling prolonged analysis of gait patterns. This approach, coupled with sophisticated algorithms for data analysis, significantly enhance the system's ability to objectively assess intoxication level, thereby substantially reducing reliance on human intervention. The system / method such as disclosed herein may employ ambient light sensors 868 and face tracking technologies to dynamically adjust screen lighting based on the user’s position, optimising conditions for accurate pupil and iris analysis. Near-infrared cameras may be used to facilitate precise tracking of pupil and iris movements, which help to provide a reliable assessment of intoxication. The inclusion of external IR illuminators 869 provides consistent lighting conditions, and thereby assist in obtaining clear readings of pupils in various environments. The system's use of Long Short-Term Memory (LSTM) networks in data training enhances its capability to recognize and learn from sequential data patterns, further improving the accuracy of its predictions. In addition, the approach described herein supports the incorporation of features like stride length and time analysis, stride velocity, upper body symmetry, swaying and staggering to distinguish between intoxication and sobriety. Advanced machine learning algorithms may be used to differentiate between dancing and walking, thereby helping to ensure accurate assessments in diverse contexts. The system 800 may have a gender segregation facility to tailor the analysis based on physiological differences, and further may have a capability to classify various physical conditions for equitable and non-discriminatory assessments. Iris tracking may be utilized for detailed intoxication level analysis, with a focus on Horizontal Gaze Nystagmus (HGN) to provide increased precision for the system's assessments. Accordingly, the limitations of existing systems are addressed by integrating a suite of sophisticated technologies for non-invasive, accurate, and efficient intoxication assessment. This approach not only enhances safety and compliance in the hospitality industry but also improves operational efficiency and customer experience. To provide a further illustration of the range of implementations that may be utilised in a system 800 for automatically dispensing alcoholic beverages, such a system may incorporate one or more of the following particular components (functionality): A) Face Tracking for Gait Measurement Termination This technology may utilise input from the camera(s) 867 in conjunction with computer vision and depth analysis to determine (inter alia) the appropriate time to stop gait analysis by calculating the person's distance from the screen. The camera(s) (together with computer vision and / or depth analysis) may also incorporate pupil tracking technology for assessing Horizontal Gaze Nystagmus (HGN), thereby aiding in determining levels of intoxication (and potentially other physical conditions). Analysis of (video) images from the camera(s) may be performed, for example, by the control software 848, in some implementations in combination with at least one ANN 818. B) Ambient Light Sensors and Face Tracking: The system 800 may use ambient light sensors 868 combined with face tracking algorithms to dynamically adjust the lighting of the screen 887 and surrounding environment based on the user’s position and distance. This adjustment of ambient light helps to enhance or optimise visual conditions for accurate pupil and iris analysis. C) Near Infrared Cameras: As discussed above, the system 800 may include one or more camera(s) 867 that are sensitive to infrared (IR) light. Such an IR camera 867 may be implemented in a camera that also provides optical imaging, or may be specialised for just IR imaging. Utilising an IR camera generally helps to enhance visibility and tracking accuracy under varied lighting conditions, which in turn facilitates precise tracking of pupil and iris, which are important to obtain accurate pupil size calculation and HGN assessment. D) External IR Illuminators: One or more external infrared (IR) illuminators 869 may be provided to ensure consistent lighting, including in the dark at nighttime. The illumination may be automatically activated or deactivated to save power, based on ambient lighting conditions. The provision of such IR lighting helps to provide consistent and clear readings of the pupils under various environmental conditions. E) Eye Tracking: Advanced eye-tracking systems may be used to focus on pupils and iris and may be implemented, for example, using one of the ANNs 818 and / or control software 847, 848. Such an eye-tracking system may tracks where the user is looking during an interaction with the system 800, and assesses for signs of nystagmus or other eye movement irregularities (which may be used in determining intoxication levels and other physical conditions). F) Iris Tracking for Intoxication Level Analysis: Iris tracking may be performed using input from camera(s) 867 and corresponding computer analysis, such as may be performed using an ANN 818, while a user 6 interacts with a screen 887. This iris tracking helps to assess intoxication level by analysing eye movements and reactions, particularly focusing on HGN. G) LSTM (Long Short-Term Memory) for Data Training: This technology is used to provide long short-term memory networks, which are a type of recurrent (artificial) neural network. The LSTM enhances the machine learning ability of a model to recognize and learn from sequential data, thereby improving prediction accuracy over time. The LSTM may be implemented on an ANN. See also Figure 5 and associated description. H) Deep Sorting Algorithms for Individual Tracking: This technology may be used to perform advanced deep learning algorithms and may be implemented, for example, by Al host software 847 and / or ANN 818 as part of the server 840. The technology may also be performed by suitable control software 848 and / or ANN 818 in a local processing unit 814, and may exploit computer vision and similar functionality. The technology may be used, inter alia, to identify and assign unique IDs to individuals in crowded spaces to track their movement towards the machine. This allows for a more prolonged analysis of gait for a given individual, thereby helping to ensure more accurate assessments and tailored interactions as individuals approach a drink dispensing machine 10. In some implementations, the system may track a single individual at a time, in other implementations, the system 800 may track multiple different individuals at a time. I) Algorithm for Averaging Analysis Results: This functionality employs a sophisticated algorithm that utilises unique IDs to check for the presence of critical markers, e.g., gait, HGN, pupil size and other physical parameters. This functionality may be performed locally in processing unit 814 or remotely on server 840 and may be used to average out all calculations of physical parameters, or give priority to specific markers for such parameters, based on the availability of the markers. J) Stride Length and Stride Time Analysis: The system 800 may train a model to discern between intoxication and sobriety based, inter alia, on stride length and time, stride velocity, upper body symmetry, swaying and staggering. Such a model may be implemented for example on an ANN 818. This monitoring of stride pattern (length and time) may be utilised in some implementations in conjunction with pose estimation technology to analyse movements and behaviour patterns, thereby complementing other parameters such as HGN and pupil size. K) Distinguishing Dancing from Walking: The system may use one or more advanced machine learning algorithms, such as may be implemented on one or more ANNs 818 (or any other suitable AI / ML platform) to differentiate between dancing and walking movements using image data from one or more cameras 867. This ability to determine whether a user 6 is dancing or walking, indicative of the gait of a user, helps to ensure accurate assessment by identifying the appropriate context of movement for analysis L) Gender Segregation for Enhanced Analysis Accuracy: The system 800 may include the capability to identify and segregate genders, such as by using one or more ANN 818 located on the processing unit 814 or remotely on a server 840 to monitor the appearance and movement of a user 6 such as by using cameras 867. Once the gender has been determined, subsequent analysis by system 800 and interaction with the user 6 may be tailored according to physiological and biomechanical differences between genders, thereby helping to enhance personalization and accuracy (such as in determining the level of any intoxication). For example, the system 800 may create (or adjust) specific models based on Gender to enhance accuracy. (In a similar approach, Height may also be assessed and then used to create and / or adjust specific models). M) Condition Classification to Avoid Discrimination: In a similar manner to gender segregation as discussed above, a classification of a user 6 may be implemented to determine various physical conditions, for example based on gait analysis, to provide equitable assessments across different potential physical factors. This classification helps to ensure that the system 800 does not discriminate between users based on their physical abilities or conditions. Various features and combinations of features are described in the following numbered paragraphs. The features may be used as appropriate in conjunction with the system and method of the appended claims. 1. An intoxication detection system 800 for automated bar environments, comprising: a processing unit 818 configured to analyze and assess visual and physiological metrics related to intoxication; and a machine learning algorithm trained to correlate anonymized user metrics with levels of intoxication. 2. The system of paragraph 1, further comprising: a camera system 867 comprising at least one camera equipped with infrared and depthsensing capabilities tailored for detection of one or more visual markers indicative of intoxication of one or more users, in which the camera(s) is / are located at one or more predetermined locations, and in which the one or more visual markers are selected from: Gait Analysis, Horizontal Gaze Nystagmus (HGN) and / or pupil dilation, or any combination thereof. 3. The system of paragraph 2, wherein the camera system, for example the at least one camera, is integrated within an interactive user interface 2 to ensure focused visual data capture from users interacting with the system. 4. The system of paragraph 2 or 3, wherein the camera system is configured in use to employ noise reduction algorithms for filtering out background activities and isolating a user's metrics even In crowded settings. 5. The system of any preceding paragraph, wherein the processing unit is configured in use to leverage convolutional neural networks (CNN) 818 for visual data interpretation and long short-term memory (LSTM) networks and to analyze sequential patterns in the one or more visual indicators. 6. The system of any preceding paragraph, further comprising a Reinforcement Learning from Human Feedback (RLHF) mechanism configured in use to allow the system to refine its detection capabilities based on real-world feedback and data. 7. The system of paragraph 1 or 2, in which the processing unit is configured in use for collaborative operation with external systems. 8. The system of any preceding paragraph, further comprising at least one server840 and at least one encrypted communication pathway configured in use to enhance data privacy and security. 9. The system of any preceding claim, further comprising a secondary user interface configured in use to allow one or more of: (i) human supervisors to oversee system data in real-time, (ii) human supervisors to contribute to RLHF inputs; and / or (iii) human users to inform of on-site alcohol service decisions.

Claims

1. A system for dispensing an alcoholic drink to a user, the system comprising:at least one artificial neural network (ANN) configured to receive, as input data, image data relating to: (i) the gait of the user; (ii) the pupil size of the user; and (iii) the Horizontal Gaze Nystagmus (HGN) of the user, wherein the at least one artificial neural network is further configured to process the input data to produce an output which indicates whether or not the user is intoxicated, and if so, the system is configured to prevent dispensing of an alcoholic drink to the user.

2. The system of claim 1, further comprising one or more imaging devices configured to obtain the image data relating to: (i) the gait of the user; (ii) the pupil size of the user; and (iii) the Horizontal Gaze Nystagmus (HGN) of the user.

3. The system of claim 2, wherein the system includes a drink dispensing device and the one or more imaging devices comprise at least first and second cameras which are incorporated into or mounted adjacent to the drink dispensing device.

4. The system of claim 3, wherein the first camera is configured to provide images for determining pupil size and HGN, optionally wherein the first camera includes a facial imaging unit.

5. The system of claim 3 or 4, wherein the second camera is configured to provide images for determining gait, and optionally wherein the second camera has a higher positioning than the first camera.

6. The system of any preceding claim, further comprising one or more ambient sensors for assessing one or more environmental parameters local to the user, optionally wherein the one or more environmental parameters include a measurement of the light level.

7. The system of claim 6, wherein measurements from the one or more ambient sensors are provided to the at least one artificial neural network and contribute to producing an output which indicates whether or not the user is intoxicated.

8. The system of any of preceding claim, wherein the system includes a lighting device for providing infrared illumination of the user.

9. The system of any preceding claim, wherein the system is configured to anonymise the input data prior to processing by the at least one ANN.

10. The system of any preceding claim, wherein the at least one ANN comprises a first ANN for analysing the gait of the user from the image data, a second ANN for analysing the pupil size of theuser from the image data, and a third ANN for analysing the Horizontal Gaze Nystagmus (HGN) of the user from the image data,11. The system of claim 10, wherein each of the first ANN, the second ANN, and the third ANN makes a respective determination of whether the user is intoxicated, and optionally wherein the respective determinations of the first, second and third ANNs are combined to produce an overall determination of whether the user is intoxicated.

12. The system of claim 11, wherein the first, second and third ANNs are allocated respective weightings for producing the overall determination, optionally wherein the weighting is determined based on data from a camera and / or sensor local to the user.

13. The system of any preceding claim, wherein the system is configured to determine height and / or gender of the user, and optionally wherein the height and / or gender are provided as additional input.

14. The system of any preceding claim, wherein measurements of gait, pupil size and HGN are supplemented by at least one further measurement device which provides additional input comprising: a depth sensor to measure distance of the user; an infrared camera; a gaze tracking system for following view direction; and / or a timer for determining blink rate15. The system of claim 13 or 14 wherein the additional input data is used for: (i) processing by the at least one ANN; and / or (ii) using the additional input data to adjust algorithms of the ANN16. The system of any preceding claim, wherein gait is determined based on landmarks on the body of the user, and / or wherein gait is used to determine one or more features including stride length and time analysis, stride velocity, upper body symmetry, swaying and staggering to distinguish between intoxication and sobriety.

17. The system of any preceding claim, the system being configured in use to leverage convolutional neural networks (CNN) for visual data interpretation and long short-term memory (LSTM) networks to analyze sequential patterns in one or more visual indicators.

18. The system of any preceding claim, further comprising a Reinforcement Learning from Human Feedback (RLHF) mechanism configured in use to allow the system to refine its detection capabilities based on real-world feedback and data.

19. The system of any preceding claim, the system comprising:a drink dispensing device local to the user and configured to support input / output communications with the user;at least one processing unit hosting the at least one artificial neural network to produce the output which indicates whether or not the user is intoxicated, and if so, the system is configured to prevent dispensing of an alcoholic drink to the user by the drink dispensing device;optionally wherein the at least one processing unit is located at the drink dispensing device and / or at a remove server system.

20. The system of any preceding claim, the system further comprising a control console to provide human monitoring of the system, wherein a human operating the control console is able to overturn a decision of the least one ANN that a user is intoxicated, thereby allowing the user to obtain an alcoholic drink from the drink dispensing device.

21. The system of claim 20, wherein the decision by the human is applied to the at least one ANN as additional training data.

22. A method for operating the system of any preceding claim for dispensing drinks.

23. A method of operating a drink dispensing system to process a request by a user for analcoholic drink, the method comprising:receiving image data of the user;using the image data for measuring (i) the gait of the user; (ii) the pupil size of the user; and (iii) the Horizontal Gaze Nystagmus (HGN) of the user;processing the input by at least one artificial neural network (ANN) for modelling (i) the gait of the user; (ii) the pupil size of the user; and (iii) the Horizontal Gaze Nystagmus (HGN) of the user;generating an output from the at least one ANN to indicate whether the user is intoxicated; andinstructing the drink dispensing system to deny or accept the request by the user according to whether or not the user is indicated as intoxicated.

24. An intoxication detection system for automated bar environments, comprising: a processing unit configured to analyze and assess visual and physiological metrics related to intoxication; anda machine learning algorithm trained to correlate anonymized user metrics with levels of intoxication.optionally wherein the intoxication detection system is combined into or configured to operate with the system of any of claims 1 to 23.

25. The intoxication detection system of claim 24 configured to perform the method of claim 22 or 23.34

Images