Directing an airflow towards a human

Abstract

An appliance comprises a first unit 102 configured to direct an airflow 105, a thermal imaging device 104, and a control unit. The control unit is configured to obtain data representing at least a por

Description

Field of the Invention The present invention relates to an appliance, and more particularly to an appliance that directs and controls an airflow towards a human. Background of the Invention Appliances, such as fans and air purifiers, can direct an airflow towards a human user, for example in order to cool the user and / or deliver purified air to the user. A known such appliance has a remote control which a user can operate in order to change the direction in which the airflow is expelled from the appliance. By appropriate operation of the remote control, the user can control the appliance to expel the airflow towards where they are located. As such, the user can control the appliance to direct the airflow towards them. Summary of the Invention According to a first aspect of the invention, there is provided an appliance comprising: a first unit configured to direct an airflow; a thermal imaging device; and a control unit; wherein the control unit is configured to: obtain data representing at least a portion of a thermal image captured by the thermal imaging device; determine, based on the obtained data, whether the at least a portion of the thermal image represents a human; and in response to a determination that the at least a portion of the thermal image represents a human, control the first unit to direct the airflow towards the human. This allows for an airflow to be directed towards a user automatically, without the user having to input control commands to direct the airflow towards them, and without the user needing to carry any specific hardware that indicates their location to the appliance. An airflow directed towards the user is therefore provided more conveniently. The use of the thermal imaging device allows that the airflow may be automatically directed towards a user without needing to capture a photographic image of the user. This allows the airflow to be directed without capturing or processing information that personally identifies any particular user. Further, this allows for a computationally efficient way to automatically direct airflow towards a user. Specifically, thermal images include information indicative of a temperature of the objects represented thereby. Since humans are typically of a different temperature to their surroundings, detecting that a thermal image represents a human can be more computationally efficient. This may be, for example, as compared to determining whether a photographic image, for example captured by a digital camera, represents a human, which requires intensive image processing to separate the human form from background included in the photographic image. Computationally efficient automatic direction of an airflow towards a user is therefore provided. Computational efficiency may be of particular importance to appliances, which may have limited processing power. Further, determining whether the thermal image represents a human, and directing the airflow towards the user in response to a determination that the thermal image represents a human, allows for a reliable direction of airflow towards the user. Specifically, by determining that the thermal image represents a human and directing the airflow in response, the airflow can be reliably directed to a human user, for example as opposed to non-human objects that may be in the environment, such as a hot coffee cup or non-human animals such as dogs, for example. Reliable automatic direction of an airflow towards a user may therefore be provided. The control unit may be configured to input the data into a trained machine learning, ML, model, the ML model having been trained to, based on an input of particular thermal image data representing a particular thermal image, output a determination of whether the particular thermal image represents a human; and obtain an output from the trained ML model thereby to determine whether the at least a portion of the thermal image represents a human. Using a trained ML model allows for the determination of whether the thermal image represents a human to be made flexibly with respect to different environments in which the appliance may be deployed. This is as compared to, for example, human presence detection j using hard-coded rules, which may not be able to produce accurate results in situations which were not explicitly considered during development. The ML model may comprise a human presence classifier, the human presence classifier having been trained to, based on an input of particular thermal image data representing a particular thermal image, classify the particular thermal image according to whether a human is present in the particular thermal image; and the output may comprise the classification. Applying a human presence classifier to the thermal image may allow for an efficient determination of human presence in the thermal image. This may be as compared to, for example, object detection whereby individual objects represented in the image are first detected and then classified. The efficient determination of human presence in the thermal image may be particularly suited to an appliance, which may have limited processing power. In examples, the human presence classification can be used a pre-condition to further processing being performed on the image, such as detection of a location of a human represented in the image. This may allow for the processing performed on each image to be minimized, and hence for efficient operation. In examples, the detection of the presence of a human in the thermal image may itself be sufficient to effectively locate the human and hence control the airflow to be directed to the human (e.g., where the human is already located at the centre of the thermal image). In other examples, the airflow may be more precisely directed by determining the location within the thermal image at which the human is represented. The human presence classifier of the ML model may have been trained based on training data comprising a plurality of sets of thermal imaging data each representing a thermal image, each set of thermal imaging data being labelled according to whether the represented thermal image represents a human. The ML model having been trained in this way may allow the ML model to determine the presence of a human in a thermal image in a variety of situations and deployments of the appliance, without any assumptions being made as to how humans are represented in the thermal imaging data. This may allow for flexible deployment of the appliance. In examples, the ML model may have been trained by supervised learning on the basis of this training data. The control unit may be configured to, prior to determining whether the at least a portion of the thermal image represents a human: interpolate the obtained data to increase the resolution of the at least a portion of the thermal image represented by the obtained data. Increasing the resolution of the thermal image may allow for the human presence to be determined more reliably, specifically by increasing the number of data points on the basis of which the human presence is determined. Accordingly, increasing the resolution by interpolation may allow for the determination to be performed more reliably, whilst still preserving the anonymity of the human by capturing a relatively low-resolution thermal image of the human. The control unit may be configured to, prior to determining whether the at least a portion of the thermal image represents a human: apply a smoothing operation to the interpolated data to smooth the at least a portion of the thermal image represented by the obtained data. This may mitigate the effects of any noise that may have been introduced through the interpolation. This may further improve the reliability of the determination as to whether a thermal image represents a human. Further, smoothing the interpolated image may allow for improved image segmentation. In examples, segmentation may be used to determine the location at which a human is represented in the thermal image. Accordingly, smoothing the interpolated image may, in turn, improve the reliability of the determination of the location at which a human is represented in a thermal image. This, in turn, may improve the precision with which the airflow may be directed towards a user. The control unit may be configured to, prior to determining whether the at least a portion of the thermal image represents a human: apply a pre-processing operation to the obtained data. In embodiments, the pre-processing operation may be configured to: interpolate the obtained data to increase the resolution of the at least a portion of the thermal image represented by the obtained data; and apply a smoothing operation to the interpolated data to smooth the at least a portion of the thermal image represented by the obtained data. The control unit may be configured to: determine, based on the obtained data, a location within the at least a portion of the thermal image at which a human is represented; and control the first unit to direct the airflow towards a location corresponding to the determined location. This may allow for the airflow to be directed more precisely towards a user. For example, in the case that the thermal imaging device has a relatively wide field of view, the location of a human within that field of view can be determined, and the airflow accordingly more precisely directed towards the user. The determination of the location within the at least a portion of the thermal image may be responsive to a determination that the at least a portion of the thermal image represents a human. This may provide an efficient way to process the obtained data. For example, this may ensure that the processing associated with determining the location within a thermal image at which a person is represented (such as processing related to segmentation of the thermal image and human centroid determination) is only applied if it is determined that the image represents a human. This may be as compared, for example, to object detection techniques where such processing may be performed on all images even if they do not include a human. In some embodiments, if it is determined that the at least a portion of the thermal image represents a human, the control unit may then determine whether the step of determining the location of the human within the thermal image is required. In some embodiments, for example where the human is already located at the centre of the thermal image, the control unit may determine that determining the location of the human within the thermal image is not required. The control unit may be configured to: determine, based on the obtained data, a location of a centroid of a human represented by the thermal image; and determine the location of the centroid as the location within the at least a portion of the thermal image at which a human is represented. The location of the centroid of the human may be determined by detecting the centre of thermal mass of the human. This may allow for the airflow to be directed at the centre of the user (e.g., the centre of thermal mass of the user), and hence for the airflow to be directed yet more precisely towards the user. The control unit may be configured to: apply a threshold to the obtained data to identify a first part of the at least a portion of the thermal image that represents a human; and determine the location of the centroid of the human based on the identified first part. The threshold may be based on a typical human body temperature range, such as 36 to 38 degrees Celsius. Applying the threshold may provide a computationally efficient way to segment the thermal image into human versus background. In turn, this may provide an efficient way to determine the centroid of the human. This may be as compared to, for example, semantic segmentation which is computationally expensive. Computationally efficient processing may be particularly important for an appliance, which may have limited computational power. In embodiments, the control unit may be configured to: apply a threshold to the obtained data to identify a first part of the at least a portion of the thermal image that represents a human; convert the first part into a segmented image; determine the location of the centroid of the human based on the identified first part; and control the first unit to direct the airflow towards a location corresponding to the location of the centroid. A segmented image may be defined as a binary image whereby each human present in the image is represented by a first value and the remainder of the image (e.g., the background) is represented by a second value. In examples, the values may be selected from 1 and 0. Converting the first part into a segmented image may make it easier to determine the boundary of the first part, and thereby make it easier to determine the location of the centroid of the human. The control unit may be configured to: determine, based on the obtained data, a distance of one or more humans, represented by the at least a portion of the thermal image, from the thermal imaging device; and control the appliance based on the determined distance. This allows for the appliance to be controlled based on the determined distance of one or more humans from the thermal imaging device and hence, in turn, a distance of one or more humans from the appliance. In examples, the control of the appliance may comprise determining to which human to the airflow is to be directed, for example directing the airflow to the closest human. In examples, the control of the appliance may comprise control of the speed of the generated airflow, for example increasing the speed of the airflow the further the human is from the appliance. In any case, controlling the appliance based on the determined distance may allow for a more flexible operation of the appliance. The control unit may be configured to: determine, based on the obtained data, the distance of a particular human from the thermal imaging device based on a size of a representation of the particular human in the thermal image. This may allow for the distance to be determined based on the thermal imaging data itself. This allows for a relatively inexpensive determination of distance. This may be as compared to, for example, using depth data such as from a time-of-flight based device. The control unit may be configured to: determine the distance of a first human and a second human, represented by the at least a portion of the thermal image, from the thermal imaging device; and control the first unit to direct the airflow to the first human in preference to the second human on the basis of the distance determined for the first human and the second human. This may allow for the airflow to be directed to a particular human based on distance from the appliance. For example, the airflow may be directed to the human closest to the appliance. This may allow that, for example, even if other humans are present in the background, the airflow will still be directed towards a user. This, in turn, may allow for the airflow to be directed more reliably towards the user of the appliance. The appliance may comprise an airflow generator configured to generate the airflow, and the control unit may be configured to: determine, based on the determined distance, a particular speed at which airflow is to be controlled to be generated; and control the airflow generator to generate an airflow at the particular speed. This may allow for the airflow experienced by the user to be consistent, even when the user is at different distances from the appliance. For example, the larger the distance of the user from the appliance, the faster the speed at which the airflow may be generated. If, however, a user moves closer to the appliance, the speed at which the airflow is generated may be reduced, for example in order to keep consistent the strength of airflow experienced by the user. The control unit may be configured to: determine, based on the obtained data, a temperature of a human represented by the thermal image data; and control the operation of the appliance based on the determined temperature. This may allow for further automated operation of the appliance on the basis of a temperature of a user (rather than the temperature of the surrounding room). For example, if it is determined that the user is hot, a speed at which an airflow is generated by the appliance may be increased and / or an airflow cooling function of the appliance may be activated or otherwise controlled to reduce a temperature of the generated airflow. As another example, if it is determined that the user is cold, an airflow heating function of the appliance may be activated or otherwise controlled to increase the temperature of the generated airflow. As such, the control unit may allow for the first unit to provide distance-based fan control, and also intelligent thermal comfort. This may allow for the airflow experienced by the user to be personalised and customised based on the user’s personal preferences. For example, the user can be directly heated or cooled to their preferred temperature from a determined (initial or starting) temperature, based on the obtained data, rather than to a preset temperature based on the temperature of the surrounding room. The first unit may comprise a first part from which the airflow is expelled and a second part, the first unit may be configured to direct the airflow by orientating the first part relative to a second part, and the first part may comprise the thermal imaging device. The thermal imaging device being part of the first part which is orientated to direct the airflow may allow for thermal images to be captured over an area larger than the field of view of the thermal imaging device, but without having to provide a separate means by which to orientate the thermal imaging device. This may save the expense and complication of providing a separate means by which to orientate the thermal imaging device. Further, this may provide for simpler control of the first unit. For example, this may allow for a consistent mapping of the airflow direction onto a viewing direction of the thermal imaging device, independent of the airflow direction. This may provide for simpler control as compared to, for example, a system in which the direction of the airflow is independent of the viewing direction of the thermal imaging device, and hence for which a mapping of the viewing direction of the thermal imaging device onto the airflow direction may have to be repeatedly determined. Efficient control may therefore be provided. In embodiments, the thermal imaging device may be integrated into, or otherwise attached to, the first part, such that the thermal imaging device is arranged to move simultaneously with the first part relative to the second part. This arrangement may allow a viewing direction of the thermal imaging device to be aligned with a direction in which the first part is configured to expel the airflow. This may allow for the control unit to determine, by determining whether the obtained data represents a human, whether a human is at a location at which the airflow is being directed. This may allow, for example, for the airflow to be directed towards a user whilst simultaneously detecting and tracking the user. This may provide for airflow to be consistently delivered to the user. Furthermore, aligning the viewing direction of the thermal imaging device (e.g., towards the centre of the user) with a direction in which the first part is configured to expel the airflow, may allow the control unit to determine that the airflow is being delivered directly to the user (e.g., delivered towards the centre of the user). The control unit may be configured to: determine, based on the obtained data, a location within the at least a portion of the thermal image at which a human is represented; determine an offset of the determined location from a central region of the at least a portion of the thermal image; and control the first unit to change the orientation of the first part to reduce the offset. This may provide a computationally efficient way to align the viewing direction of the thermal imaging device (and hence the direction in which the airflow is expelled) with the location of the user. This allows for tracking of the location of the user and that, even in cases where the user moves relative to the appliance, the appliance can direct the airflow towards the user. The control unit may be configured to: determine, based on a first set of obtained data representing at least a portion of a first thermal image captured by the thermal imaging device at a first time, a first location within the at least a portion of the first thermal image at which a human is represented; determine, based on a second set of obtained data representing at least a portion of a second thermal image captured by the thermal imaging device at a second time, a second location within the at least a portion of the second thermal image at which a human is represented; determine, based on the first location and the second location, a direction of travel of the human; and control the first unit to rotate the first part relative to the second part according to the determined direction of travel. This allows for tracking of the user (and hence for the airflow to be directed towards the user) even in cases where the user moves so as to leave the field of view of the thermal imaging device, for example, in the case the user moves faster than the first part of the first unit can be re-orientated. In examples, the control unit may control the first unit to rotate the first part relative to the second part in the determined direction of travel until the control unit determines again that a human is represented by a captured thermal image. This may allow the airflow to be continuously delivered towards the user, for example towards the centre of the user, even in cases where the user moves relative to the appliance or where the user moves so as to leave the field of view of the thermal imaging device. The control unit may be configured to: control the first unit to rotate the first part relative to the second part; in response to the determination that the at least a portion of the thermal image represents a human, control the first unit to stop rotating the first part relative to the second part. This allows that a user need not necessarily first walk into the field of view of the thermal imaging device in order for the appliance to direct the airflow towards the user. Instead, the appliance may rotate the first part through certain rotation window in order to scan for a human to which to direct the airflow. Once the control unit determines that a captured thermal image represents a human, the control unit may cease rotating the first part. In this way, the airflow may be directed towards a user even though the field of view of the thermal imaging device may not initially incorporate the location of the user. The convenience of operation of the appliance may therefore be further improved. Optional features of aspects of the present invention may be equally applied to other aspects of the present invention, where appropriate. Brief Description of the Drawings Figure lisa schematic diagram illustrating a perspective view of an appliance according to an example; Figure 2 is a schematic diagram illustrating a cross section of the example appliance of Figure 2; Figure 3 is a schematic flow diagram illustrating a process according to an example; Figure 4 is a schematic flow diagram illustrating a process according to another example; Figure 5 is a schematic diagram illustrating a thermal image according to an example; Figure 6 is a schematic diagram illustrating four thermal images with different resolutions, according to an example; Figure 7 is a schematic diagram illustrating a machine learning model, according to an example; Figure 8 is a schematic diagram illustrating nine thermal images of a training data set, according to an example; Figure 9 is a schematic diagram illustrating three thermal images at various stages of a segmentation process, according to an example; Figure 10 is a schematic diagram illustrating an array of sensor of a thermal imaging device, according to an example; and Figure 11 is a schematic diagram illustrating a control unit, according to an example. It is noted that each of Figures 1, 2 and 10 show a coordinate system x, y, z. This coordinate system is common to each of Figures 1, 2 and 10. Detailed Description of the Invention Referring to Figures 1 and 2, there is illustrated an appliance 100 according to an example. The appliance 100 is for directing an airflow 105 towards a human user 240. For example, the appliance 100 may be a personal fan, heater, and / or air purifier. For example, the appliance 100 may be for use in an indoor environment, such as in a room. The appliance may be referred to as an environmental care appliance 100, in that the appliance 100 is for affecting an environment, such as an indoor environment, in which a human user may be situated. The appliance 100 comprises a first unit 102 configured to direct the airflow 105. The first unit 102 comprises a first part 114 and a second part 116. The first part 114 comprises an airflow outlet 112 from which the airflow 105 is expelled from the appliance 100. The second part 116 provides a base via which the appliance 100 is supported on a flat surface 231 such as a floor or a table. The first part 114 is configured to rotate 120 relative to the second part 116 about a rotational axis 122 of the appliance 100. The first part 114 of the first unit 102 comprises an air inlet 118 through which air enters the appliance 100, an airflow generator 234 configured to generate the airflow 105, and the airflow outlet 112 from which the generated airflow 105 is expelled from the appliance 100. The airflow generator 234 may comprise an electric motor (not shown) configured to drive an impeller (not shown). Operation of the airflow generator 234 is controlled by a control unit 232. In operation, the airflow generator 234 generates the airflow 105 by drawing air in through the air inlet 118 and pushing the drawn in air towards the airflow outlet 112. In examples, the appliance may comprise a filter (not shown) configured to filter air drawn into the appliance 100 through the air inlet 118. In such examples, the filter (not shown) may remove impurities from the air and, as such, the appliance 100 may be referred to as an air purifier. In examples, the appliance 100 may comprise a humidifier (not shown) configured to humidify air drawn into the appliance through the air inlet 118. In such examples, the appliance 100 may be referred to an air humidifier. The airflow outlet 112 of the first part 114 is configured to expel the airflow 105 in a particular direction 106. Specifically, the outlet 112 is generally annular in shape and acts to expel the airflow 115 in a direction 106 generally perpendicular to the plane in which the annular outlet 112 is defined. The first part 114 comprises a bore 111 defined inwardly of the annular outlet 112. In some cases, the airflow expelled from the annular outlet 112 may entrain surrounding air via the bore 111, in which case the directed airflow 105 may comprise both the airflow expelled from the outlet 112 and the flow of entrained air. In either case, the airflow 105 flows in the particular direction 106. As mentioned, the first unit 102 is configured to direct the airflow 105. The first unit 102 is configured to direct the airflow 105 by orientating the first part 116 relative to the second part 114. That is, a change in the orientation of the first part 116 relative to the second part 114 changes the direction in which the airflow 105 is expelled from the appliance. More specifically, a change in the rotational orientation of the first part 116 relative to the second part 114 about the rotational axis 122 of the appliance 100 changes the direction in which the airflow 105 is expelled from the appliance 100. The first unit 102 comprises a motor assembly 230 configured rotate 120 the first part 116 relative to the second part 114 about the rotational axis 122. The control unit 232 controls the first unit 102 to direct the airflow 105 in a particular direction, that is, a direction having a particular rotational orientation about the rotational axis 122. Specifically, the control 232 controls the motor assembly 230 to rotate the first part 116 relative to the second part 114 so that the first part 116 is in a particular rotational orientation relative to the second part 114, thereby to control the airflow 115 to be provided in a direction having a particular rotational orientation about the rotational axis 122. By controlling the airflow generator 234 and the motor assembly 230, the control unit 232 controls both the speed and direction (that is, the rotational orientation about the rotational axis 122) of the airflow 105. The appliance 100 comprises a thermal imaging device 104. More particularly, the first part 114 of the first unit 102 of the appliance 100 comprises the thermal imaging device 104. The thermal imaging device 104 is configured to capture one or more thermal images of space surrounding the appliance 100. The thermal imaging device 104 may capture a series of thermal images, also referred to herein as frames. The thermal imaging device 104 may comprise a 2-dimensional array of non-contact thermal sensors, such as Infrared sensors. For example, the thermal imaging device 104 may comprise a 2-dimensional array of thermopile sensors configured to sense incident IR flux. The thermal imaging device 104 may be a relatively low-resolution thermal imaging device 104. That is, the thermal images captured by the thermal imaging device 104 may be of relatively low resolution. This may be as compared to, for example, typical photographic digital images which typically have resolutions on the order of thousands of pixels. For example, the thermal imaging device 104 may be a 64-pixel thermal imaging device 104, for example with an 8x8 array of non-contact thermal sensors. That is, the thermal images captured by the thermal imaging device 104 may have a size of 64 pixels, and dimensions of 8 pixels by 8 pixels. In examples, the thermal imaging device 104 may be a 10080 pixel thermal imaging device, for example with a 120x84 array of non-contact thermal sensors. In examples, the thermal imaging device 104 may have a resolution in the range 64 to 10080. Other example resolutions and array sizes are possible. In any case, the thermal imaging device 104 is configured to capture thermal images. The thermal imaging device 104 is configured to output data representing the captured thermal images to the control unit 232. For example, data representing a thermal image may comprise a pixel value for each pixel of the thermal image. The thermal imaging device 104 is configured to capture thermal images along a viewing direction 108, and with a particular field of view 110. The viewing direction 108 may be thought of as, for example, the focal axis 108 of the thermal imaging device 104. A width of the field of view 100 may, for example, represent 60 degrees about the rotational axis 122 of the appliance 100. The viewing direction 108 of the thermal imaging device 104 is substantially the same as the direction 106 in which the airflow 105 is directed by the first unit 102. That is, the rotational orientation of the viewing direction 108 about the rotational axis 122 of the appliance 100 is the same as that of the direction in which the first unit directs the airflow 105. At any given time, the region of space of which the thermal imaging device 104 captures thermal images may be substantially the same as (or at least include) the region of space towards which the airflow 105 is directed. The control unit 232 is configured to control the direction in which the airflow 105 is provided by the first unit 102 based on data representing at least a portion of a thermal image captured by the thermal imaging device 104. More specifically, the control unit 232 is configured to control the first unit 102 to orientate the first part 114 in a particular rotational orientational relative to the second part 116 about the rotational axis 122 of the appliance 100, thereby to direct the airflow 105 in a particular direction having a particular rotational orientation about the rotational axis 122. The appliance 100 comprises a power supply 115 configured to power components of the appliance 100, for example the control unit 232, the airflow generator 234, the motor assembly 230 and / or the thermal imaging device 104. A broad overview of the control processes that the control unit 232 is configured to implement is illustrated in Figure 3. This process may be a method performed by the control unit 232. Referring to Figure 3, in step 302, the control unit 232 obtains data representing at least a portion of a thermal image captured by the thermal imaging device 104. For example, the control unit 232 may obtain data representing the whole of a thermal image captured by the thermal imaging device 104, or data representing a portion of a thermal image captured by the thermal image device. As mentioned, the data may comprise pixel values for each pixel of the thermal image or portion thereof. An example of a thermal image 500 is shown in Figure 5 (descried in more detail hereinafter). Returning to Figure 3, in step 304, the control unit 232 determines, based on the obtained data, whether the at least a portion of the thermal image represents a human 240. For example, a thermal image represents a human when the thermal image is captured by the thermal imaging device 104 at a time when there is a human located in a region of space encompassed by the field of view 100 of the thermal imaging device 104. On the other hand, a thermal image does not represent a human when the thermal image is captured by the thermal imaging device 104 at a time when there is no human located in a region of space encompassed by the field of view 100 of the thermal imaging device 104. This includes where there are relatively hot or otherwise IR bright objects (such as hot coffee cups, laptop computers, or windows through which sunlight is shining) in the region of space encompassed by the field of view 100 of the thermal imaging device 104, but which nevertheless are not human. In examples, the determination as to whether the thermal image represents a human comprises using a machine learning, ML, model trained to classify thermal images according to whether a human is present in the thermal image (described in more detail hereinafter). If the control unit 232 determines that the at least a portion of the thermal image represents a human (YES in step 304), then the control unit 232 carries out step 306. However, if the control unit 232 determines that the at least a portion of the thermal image does not represent a human (that is, there is no human represented in the at least a portion of the thermal image; NO in step 304), then the control unit 232 may return to step 302 and obtain data representing another thermal image, for example the next thermal image captured by the thermal imaging device 104. In response to a determination that the at least a portion of the thermal image represents a human (YES in step 304), the control unit 232, in step 306, controls the first unit 102 to direct the airflow 105 towards the human 204. Controlling the first unit 102 to direct the airflow 105 towards the human 204 may be implemented in one or more or a combination of ways. In some examples, the control comprises controlling the first unit 102 to stop rotation of the first part 114 relative to the second part 116. For example, the first unit 102 may be configured to rotate the first part 114 relative to the second part 116 so as to pan the field of view 110 of the thermal imaging device 104 about the axis 122 of the appliance. For example, this may be performed as part of an initialisation sequence to determine the direction(s) in which one or more humans are located, relative to the appliance 100. As the field of view 100 is panned, the thermal imaging device 104 may capture thermal images. The control unit 232 may analyse each image as it is captured, for example in real-time or near-real time, to determine whether the image represents a human. When it is determined that a thermal image does not represent a human, the controller 232 may control or otherwise allow the pan to continue, for example by not sending a control command to the first unit 102. However, when it is determined that a thermal image represents a human, the controller 232 may control the first unit 102 to stop or cease the pan (that is, stop or cease the rotation of the first part 114 relative to the second part 116), for example by sending a control command to the first unit 102 to cease the pan. As such, since the viewing direction 105 of the thermal imaging device 104 and the direction of the airflow 105 output from the appliance 100 are substantially the same, stopping the pan when a human is detected in a thermal image captured by the thermal imaging device 104 will cause the airflow 105 to be directed towards the human. In some examples, the control comprises controlling the first unit 102 to direct the airflow 105 towards a particular location. For example, the control may comprise controlling the first unit 102 to rotate the first part 114 relative to the second part 116, for example by a particular amount in a particular direction, and / or so that the first part 114 has a particular rotational orientation relative to the second part 116. For example, the control unit 232 may determine based on the obtained data, a location within the at least a portion of the thermal image at which a human is represented, and control the first unit 102 to direct the airflow towards a location corresponding to the determined location. For example, a human may be present at an edge of the field of view 110 of the thermal imaging device 104 and hence may be represented towards an edge of a thermal image captured by the thermal imaging device. In this case, the control unit 232 may control the first unit 102 to rotate the first part 114 so that the field of view 100 is more centred on the human. This, in turn, will cause the airflow 106 can be more precisely directed towards the human. In some examples, the control comprises controlling the first unit 102 to rotate the first part 114 relative to the second part 116 in a particular rotational direction. For example, the control unit 232 may determine a direction of travel of a human relative to the appliance 100, and control the first unit 102 to rotate the first part 114 relative to the second part 116 according to the determined direction of travel. This may help ensure that a moving human can be tracked, and hence for the airflow to be directed towards the human. Referring to Figure 4, there is illustrated a flow diagram of a control process that is performed by the control unit 232, according to specific examples. Certain steps of the control process will be described in more detail with reference to certain other Figures, as indicated in more detail hereinafter. The process starts at step 402 where the control unit 232 obtains data representing at least a portion of a thermal image captured by the thermal imaging device 104. This may be the same or similar to step 302 of Figure 3 described above. Referring to Figure 5, there is illustrated a thermal image 500 according to an example. The thermal image 500 is a two-dimensional, 8 by 8 array of pixels 502. Each pixel has a value, represented in Figure 5 by the lightness or darkness of the respective pixel. The value for each pixel may, for example, represent the intensity or flux of infrared radiation at a corresponding sensor of the two-dimensional array of sensors that the thermal imaging device 104 comprises. The data representing the thermal image 500 may comprise the value of each pixel and the location of each pixel in the array. In examples, the location of the pixel in the array may be implicit in the ordering of the pixel values. In the example illustrated in Figure 5, the thermal image 500 represents a human 504. That is, when the thermal image 500 was captured, the field of view of the thermal imaging device 104 covered a region of space where a human was located. The human 504 represented in the thermal image 500 can be seen as an area of lighter pixels, representing that the human 504 is of a higher temperature than their surroundings. The human is not personally identifiable by the thermal image 500. The control unit 232 may obtain data representing the whole thermal image 500, or a portion of the thermal image 500, for example a cropped portion of the thermal image 500. The control unit 232 analyses the data representing the obtained at least a portion of the thermal image 500 to determine whether the thermal image 500 represents a human. Returning to Figure 4, once the control unit 232 obtains data representing the at least a portion of the thermal image 500, the control unit 232, in step 404 begins pre-processing of the thermal image 500. In particular, in step 404 the control unit 232 interpolates the obtained data to increase the resolution of the at least a portion of the thermal image 500 represented by the obtained data. For example, the control unit 232 may define a two-dimensional pixel grid of increased resolution and determine the pixel value for each pixel of the grid by interpolating the pixel values of the thermal image 500. For example, the control unit 232 may apply bicubic interpolation to the data to increase the resolution of the image represented thereby. Interpolation to different resolutions may be used. For example, referring to Figure 6, there is illustrated a first thermal image 500 which has an 8 by 8 pixel resolution, as per Figure 5. A second thermal image 602 in Figure 6 represents the results of applying bi-cubic interpolation to the first thermal image 500 to increase the resolution to 32 by 32 pixels. A third thermal image 604 in Figure 6 represents the results of applying bicubic interpolation to the first thermal image 500 to increase the resolution to 64 by 64 pixels. A fourth thermal image 606 in Figure 6 represents the results of applying bi-cubic interpolation to the first thermal image 600 to increase the resolution to 128 by 128 pixels. The interpolation may improve the accuracy and / or reliability with which a machine learning model can classify the thermal image according to whether the thermal image represents a human or not. Returning to Figure 4, following the interpolation, the control unit 238, in step 406, applies a smoothing operation to the interpolated data to smooth the at least a portion of the thermal image represented by the obtained data. For example, the interpolated data may represent any one of the second 602, third 604, or fourth 606 thermal images in Figure 6. The smoothing operation may comprise convolution of the data representing the interpolated image 602, 604, 606 with a Gaussian kernel, for example a two-dimensional isotropic Gaussian kernel. This smoothing operation may reduce artefacts of the interpolation, such as bleeding of pixel value from the outline of the representation of the human. This may, in turn, improve the reliability with which the ML model can classify the thermal image according to whether the image represents a human. Further, this may, improve the accuracy of image segmentation, and hence human centroid location determination, described in more detail below. Following the smoothing in step 406, the control unit 232, in step 408, determines whether the (interpolated and smoothed) thermal image represents a human. In some examples, this step may be performed using a trained machine learning, ML, model. Referring to Figure 7, there is illustrated an example machine learning model 704, having an input 702 and an output 706. The ML model has been trained to, based on an input 702 of particular thermal image data representing a particular thermal image, output 706 a determination of whether the particular thermal image represents a human. The control unit 232 may be configured to input the data representing the at least a portion of the (interpolated and smoothed) thermal image into the trained ML model 704 and obtain an output 706 from the trained ML model 704 thereby to determine whether the at least a portion of the thermal image represents a human. In this example, the ML model 704 comprises a human presence classifier 704. The human presence classifier has been trained to, based on an input 702 of particular thermal image data representing a particular thermal image, classify the particular thermal image according to whether a human is present in the particular thermal image. In this case, the output 706 of the human presence classifier 704 comprises the classification. For example, the human presence classifier may be implemented by a support vector machine. Other suitably trained classifiers, such as logistic regression, may be used. The human presence classifier for the ML model may be trained, for example, using supervised learning. For example, the human presence classifier may have been trained based on training data, the training data comprising a plurality of sets of thermal imaging data each representing a thermal image, each set of thermal imaging data being labelled according to whether the represented thermal image represents a human. The classification labels of the training data set may be determined, for example, by a person confirming that a human is (or is not) present in the respective thermal images. For example, this may be by a person confirming that a human was (or was not) present in an area covered by the field of view of the thermal imaging device and in the line of sight of the thermal imaging device when the training thermal image was captured, and labelling the training thermal image with the appropriate classification. Referring to Figure 8, there is illustrated nine thermal images 802-820 of a training data set according to an example. A human is present in each of the fifth 810, sixth 812, seventh 814 and eighth 816 thermal images. Accordingly, each of the fifth 810, sixth 812, seventh 814 and eighth 816 training thermal images may be labelled with a label indicating that the image represents a human. As can be seen, these thermal images include representations of a human in different situations, including at a desk (image 810), seated at the edge of the field of view (image 812), stood near the thermal imaging device (image 814) and at another desk (image 816). These may be referred to a positive training example, that is, examples that help the classifier to learn when a thermal image represents a human. However, each of the first 802, second 804, third 806, fourth 808, and ninth 820 training thermal images are examples of objects which do not represent a human. Accordingly, each of the first 802, second 804, third 806, fourth 808, and ninth 820 training thermal images are not typically classified and are therefore not labelled. However, in some examples, each of the first 802, second 804, third 806, fourth 808, and ninth 820 training thermal images may be labelled with a label indicating that the image does not represent a human. As can be seen, these images include representations of various scenes which do not include a human, including an empty desk (image 802), a window (image 804), a wall (image 806), another window (image 808), and another empty desk (image 820). These may be referred to as negative training examples, that is, examples that help the classifier to learn when a thermal image does not represent a human. During training based on the training data, the human presence classifier learns parameters of a model to correctly classify the sets of thermal image data of the training data according to their respective, ground-truth, labels. By training the human presence classifier on tens or hundreds or thousands of sets of thermal imaging data and their respective classification labels, the human presence classifier learns to, for an input 702 of particular thermal image data representing a particular thermal image, classify the particular thermal image according to whether a human is present in the particular thermal image. Returning to Figure 4, if, in step 408, the human presence classifier determines that the thermal image, represented by the obtained (interpolated and smoothed) thermal image data, does not represent a human (“no presence” in Figure 4), the control unit 232 returns to step 402. That is, the subsequent processing of steps 410 to 420 of Figure 4 (described hereinafter) are not performed. Instead, the control unit 232 returns to step 402 and obtains data representing another thermal image captured by the thermal imaging device 104. For example, the control unit 232 may obtain data representing the next thermal image captured by the thermal imaging device 104, and the processing steps 406 to 408 may be performed for the next thermal image. If, on the other hand, in step 408, the human presence classifier determines that the thermal image, represented by the obtained (interpolated and smoothed) thermal image data, does represent a human (“presence” in Figure 4), the control unit 232 progresses to steps 410 to 420 (described hereinafter). The control unit 232, in steps 410 to 412, determines, based on the obtained (interpolated and smoothed) thermal image data, a location within the at least a portion of the thermal image at which a human is represented. Specifically, in this example the control unit 232 determines a location of a centroid of a human represented by the thermal image (e.g., centre of thermal mass), and determines the location of the centroid as the location within the at least a portion of the thermal image at which a human is represented (described in more detail hereinafter). Then, in steps 414 to 420, the control unit 232 controls the first unit 102 to direct the airflow 105 towards a location corresponding to the determined location. Specifically, in this example, the control unit 232 determines an offset of the determined location within the at least a portion of the thermal image at which a human is represented (specifically, the location of the centroid) from a central region of the at least a portion of the thermal image, and controls the first unit 102 to change the orientation of the first part 114 to reduce the offset (described in more detail hereinafter). This causes the airflow 105 to be precisely directed towards the user. It is to be noted that in this example, the determination of the location within the at least a portion of the thermal image is responsive to the determination, in step 408, that the at least a portion of the thermal image represents a human. Accordingly, for a thermal image determined as not representing a human in step 408, steps 410 to 420 are not performed. Steps 410 to 420 of the example of Figure 4 are hereinafter described in more detail. The control unit 232, in step 410 begins post-processing of the thermal image 500. In particular, in step 410 the control unit 232 segments the thermal image. Specifically, the control unit segments parts of the (interpolated and smoothed) thermal image that represent a human. In examples, the segmentation comprises threshold segmentation (e.g., based on a typical human body temperature range). Specifically, the control unit 232 may apply a threshold to the obtained data to identify a first part of the at least a portion of the thermal image that represents (or potentially represents) a human. For example, the control unit 232 may apply a threshold to the pixel values of each pixel of the thermal image, and those pixels that have a pixel value above the threshold may be segmented as representing a human. The threshold used for a particular thermal image may be scaled with respect to the range of pixel values in any given image. For example, the threshold may be an upper quartile threshold, where pixels having a pixel value in the highest 25% of all of the pixel values of the thermal image are segmented as representing (or potentially representing) a human, and the remaining 75% are segmented as background. This may provide a threshold that is adaptive to the range of pixel values that are in any given image. In examples, the segmentation may further comprise connected component analysis, CCA. For example, CCA may be applied to the pixels segmented as representing (or potentially representing) a human by the threshold segmentation. This may identify continuous parts of the thermal image that have been segmented as representing (or potentially representing) a human. The largest continuous part of the thermal image segmented as representing (or potentially representing) a human, as determined by the CCA, may be selected as the part of the thermal image that represents the human. That is, the largest continuous part resulting from the CCA is identified as the first part of the at least a portion of the thermal image that represents the human. This may provide that smaller, non-human hot objects, such as a hot coffee cup, that are in a different location to the human, can be filtered out of the thermal image for the purposes of determining the location of the human in the thermal image. In cases where there is more than one human in the thermal image, selecting the largest continuous part of the thermal image resulting from the CCA may allow the control 232 to identify the human that is closest to the thermal imaging device, and hence to which human the airflow is to be directed. For example, the control unit 232 may be configured to direct the airflow 105 to the human closest to the thermal imaging device 104. Since the human closest to the thermal imaging device 104 will appear largest in the thermal image, the selection of the largest continuous part of the thermal image resulting from the CCA may allow the control unit 232 to identify the appropriate human to which the airflow 105 is to be directed. Following the segmentation to identify the first part of the at least a portion of the thermal image that represents a human (or, more specifically, the first part of the at least a portion of the thermal image that represents the human to which the airflow is to be directed), the control unit 232, in step 412, the location of a centroid of the human based on the identified first part. For example, the centroid of the identified first part may be taken as the centroid of the human represented by the thermal image. Accordingly, the centroid of the identified first part may be taken as the location of the representation of the human within the image. Referring to Figure 9, there are illustrated three thermal images 902, 904, 906 relating to the segmentation and centroid determination process of steps 410 and 412 of Figure 4. The first thermal image 902 is a thermal image resulting from the 128 by 128 pixel interpolation, and subsequent smoothing, of the thermal image 500 of Figures 5 and 6. The second thermal image 904 represents the thermal image resulting from applying an upper quartile threshold to the first thermal image 902. Specifically, in the second image 904, only the top 25% of pixel values remain, and other pixel values have been replaced with zero. The third thermal image 906 represents the segmentation resulting from the upper quartile threshold. Specifically, all pixels with a value greater than zero (represented in white in the third thermal image 906) are segmented as a part of the thermal image 906 that represents (or potentially represents) a human. CCA is applied to the third thermal image 906 which identifies the largest continuous part 908 of the third thermal image 906 segmented as representing a human, as the first part 908 of the thermal image that represents the human. The centroid 910 of this first part 908 is determined, and taken as the location in the thermal image at which the human is represented. Referring again to Figure 4, following the determination of the centroid 910, the control unit 232, in step 414, converts the location of the centroid in the thermal image to an angle about the rotational axis 122 of the appliance through which the first part 144 would need to be rotated in order to centre the human in the field of view 100 of the thermal imaging device, and thereby direct the airflow 105 precisely towards the human. Specifically, the conversion in step 414 comprises determining an offset of the determined location of the centroid within the thermal image from a central region, for example the centre, of the thermal image. The control unit 232 may then control the first unit 102 to change the orientation of the first part 114 to reduce the offset. That is, the control unit 232 may control the first unit 102 to rotate the first part 114 about the rotational axis 122 so as to centre the human in the field of view 100 of the thermal imaging device, and thereby direct the airflow 105 precisely towards the human. Referring to Figure 10, there is illustrated an 8 by 8 thermal image 1000 (where each pixel is white for clarity). In this case, the location of a human centroid 1002 has been determined. The distance of the centre 1005 of the thermal image from a first edge 1003 of the thermal image (that is, the distance along the x axis, that is, the horizontal axis, in the sense of Figure 10) is given by “xwidth”. The distance of the location of the human centroid 1002 from the first edge 1003 of the thermal image (that is, the distance along the x axis, that is, the horizontal axis, in the sense of Figure 10) is given by “x^centroid”. The difference between “x^centroid” and “xjwidth” is the offset, along the x axis, of the location of the centroid 1002 from the centre 1005 of the thermal image. The control unit 232 determines this offset and, using a known value of the field of view of the thermal imaging device 104 that captured the thermal image 1000, converts the offset into an angle about the rotational axis 122. Specifically, the control unit 232 converts the offset into an angle about the rotational axis 122 through which the first part 114 of the first unit 102 would need to rotate in order to align the centre of the field of view of the thermal imaging device with the human (as represented by the human centroid 1002). For example, the control unit 232 may determine the angle 0 by inputting “x width” and “xcentroid” into equation (1): @ ~ (.^centroid — “ (1) where F is the field of view of the thermal imaging device in the x direction, and N is the total width of the thermal image 1000, that is the total distance from the first edge 1003 of the thermal image to the opposing edge. In example, each of “x width”, “x centroid” and “TV” may be in units of pixels, and the angle 6 and the field of view F may be in units of degrees. For example, the thermal imaging device may have a field of view of 60 degrees. In the example of Figure 10, Nis eight pixels, “xwidth” is 4 pixels, and “xcentroid” is 6.5 pixels. Accordingly, the angle 6 may be determined as (6.5 - 4)*(60 / 8) = +18.75 degrees. This may correspond to rotating the first part 114 of the first unit 102 about the rotational axis 122 to the right in the sense of Figure 10. Returning to Figure 4, following the centroid to angle conversion in step 414, the control unit 232, in step 416, determines whether the determined angle is within a first range of angles. For example, the first range of angles may represent a tolerance within which it may be determined that the airflow 105 is currently directed towards the human and hence the rotational orientation of the first part 114 need not be changed. If it is determined in step 416 that the angle is within the first range or tolerance (“Yes” in Figure 4), then the control unit 232 determines in step 420 that the human is centred in the thermal image. In this case, the control unit does not issue a rotation command to the first unit 102, but instead returns to step 402 and obtains data representing the next thermal image captured by the thermal imaging device 104. The process of Figure 4 may then be repeated for this next thermal image. If, on the other hand, it is determined in step 416 that the angle is outside of the first range or tolerance (“No” in Figure 4), the control unit 232, in step 418, controls the first unit 102 to rotate the first part 114 about the rotational axis 122 by the determined angle. For example, in the example if Figure 10, the control unit 232 may issue a rotation command to the first unit 102 to rotate the first part 114 about the rotational axis 122 by +18.75 degrees, and the first unit 232 may respond by rotating the first part 114 about the rotational axis by +18.75 degrees (or by an amount approximating +18.75 degrees, for example to the nearest rotational step that the first unit 102 is able to rotate the first part 114). This may direct the airflow 105 precisely towards the human. After the rotation command is issued in step 418, the control unit 232 may return to step 402, and obtains data representing the next thermal image captured by the thermal imaging device 104. The process of Figure 4 ay then be repeated for this next thermal image. In this way, the user may be tracked, and the airflow directed towards the user even when the user moves around the space surrounding the appliance 100. The first range of angles (or tolerances) in step 416 helps prevent unnecessary oscillation of the first part 114 of the first unit 102 about the rotational axis 122. For example, depending on the step size of the motor assembly 230 that drives the rotation of the first part 114, a rotation command may cause the first unit 102 to slightly overshoot the rotational position of the human user. This could result in oscillation of the first part 114 about the location of the user. However, not sending the rotation command when the angle is within the first range of values can prevent the minor correction of the slight overshoot and hence prevent the unnecessary oscillation. In examples, the first range of angles may be greater than or equal to a step size of rotation of the first part 114 of the first unit 102. For example, the step size may represent the minimum amount by which first part 114 can be rotated by the motor assembly 230. The first range of angles being greater than or equal to the step size may allow that any single step cannot take the determined angle from outside of the first range on one side of the first range to outside of the first range on the opposite side of the first range. This may help ensure that the unnecessary oscillation is prevented. In examples, the control unit 232 may be configured to determine a direction of travel of the human 240 based on a plurality of captured thermal images, and control the first unit 102 to rotate the first part 114 relative to the second part 116 according to the determined direction of travel. For example, this may be used when the human moves faster than the first unit 102 can rotate the first part 114 to keep the human within the field of view 110. In this case, the control unit 232 may control the first unit 102 to rotate the first part 114 so as to pan the field of view 110 in the determined direction of travel. The pan may continue until the human presence in a captured thermal image is again determined (and the control process described above with reference to Figure 4 may again be performed). For example, the control unit 232 may determine, based on a first set of obtained data representing at least a portion of a first thermal image captured by the thermal imaging 104 device at a first time, a first location within the at least a portion of the first thermal image at which a human is represented; and determine, based on a second set of obtained data representing at least a portion of a second thermal image captured by the thermal imaging device 104 at a second time, a second location within the at least a portion of the second thermal image at which a human is represented. The direction of travel of the human may be based on the first location and the second location. For example, the first part 114 may not be rotated in between the first time and the second time. In this case, the control unit 232 may determine that, if the second location is to the right of the first location, the direction of travel is to the right, for example. In this case the control unit 232 may control the first unit 102 to pan the first part 114 to the right. If the second location is to the left of the first location, the control unit 232 may determine that the direction of travel is to the left, for example. In this case the control unit 232 may control the first unit 102 to pan the first part 114 to the left. In some examples, the control unit 232 may be configured to determine, based on the obtained thermal image data, a distance 242 of one or more humans, represented by the at least a portion of the thermal image, from the thermal imaging device 114. The control unit 232 may then control the appliance 100 or components thereof based on the determined distance. As one example, the control unit 232 may be configured to determine the distance of a first human and a second human, represented by the at least a portion of the thermal image, from the thermal imaging device 104, and control the first unit 102 to direct the airflow 105 towards the first human in preference to the second human on the basis of the distance determined for the first human and the second human. This may allow for the airflow to be directed towards the user even in the case that there are other humans in the background of the captured thermal images. As another example, the control unit 232 may be configured to determine, based on the determined distance of the human from the thermal imaging device 104, a particular speed at which the airflow 105 is to be controlled to be generated, and control the airflow generator 234 to generate an airflow at the particular speed. For example, the larger the determined distance of the human from the appliance 100, the faster the speed at which the airflow 105 may be generated, and vice versa. This may allow for the airflow experienced by the user to be consistent, even when the user is at different distances from the appliance. In some examples, the control unit 232 may be configured to determine, based on the obtained thermal image data, the distance of a particular human from the thermal imaging device 114 based on a size of a representation of the particular human in the thermal image. For example, as described above with reference to step 410 of Figure 4 and also with reference to Figure 9, where there may be two humans represented in the thermal image, the control unit 232 may apply CCA and identify the largest continuous part of the segmented thermal image as representing the human that is closest to the thermal imaging device. In other examples, the distance of a human from the thermal imaging device may be inferred directly from the size of the representation of the human in the thermal image. For example, where a human is relatively close to the thermal imaging device, the representation of the human will appear relatively large in the thermal image, and vice versa. The control unit 232 may determine the first part of a thermal image that represents a human (for example using the segmentation step 410 described above). The control unit 232 may then determine one or more dimensions of the first part, such as the height, width, and / or area of the first part. The control unit 232 may then map the determined one or more dimensions onto a distance of the human from the thermal imaging device 104. The mapping may, for example, be calibrated by appropriate tests. In cases where the thermal image represents more than one human, the control unit 232 may apply this process to each part of the thermal image determined as representing a human. In some examples, the control unit 232 may be configured to determine, based on data representing at least a portion of a thermal image captured by the thermal image device 104, a temperature of a human represented by the thermal image data. The control unit 232 may then be configured to control operation of the appliance 100 or one or more components thereof based on the determined temperature. For example, as described above with reference to Figure 5, the data representing the thermal image may comprise pixel values each representing the intensity or flux of infrared radiation at a corresponding sensor in the thermal imaging device. This value can be mapped onto a temperature of the object represented at the associated pixel. For example, this mapping may be done by the thermal imaging device 104 itself, that is, the pixel values output by the thermal imaging device 104 may be temperature values, for example in units of degrees Celsius. As another example, the control unit 232 may convert the values indicative of IR flux or intensity into a temperature, by suitable calibration. In either case, the control unit 232 may be configured to determine a temperature of the human represented in a thermal image based on the data representing the thermal image. In examples, the control unit 232 may identify the first part of the thermal image that represents a human (for example using segmentation and CCA as described above). The control unit 232 may then determine a temperature of the human represented by the thermal image based on the thermal image data corresponding to the first part of the thermal image. For example, the control unit 232 may take the highest temperature indicated in the first part of the thermal image, or an average of the temperatures indicated in the first part of the thermal image. As mentioned, the control unit 232 may be configured to control operation of the appliance 100 or one or more components thereof based on the determined temperature. For example, the control unit 232 may control the speed at which the airflow is generated by the airflow generator 234 based on the determined temperature. For example, if it is determined that the human is relatively hot (for example, the determined temperature is above a first threshold), the speed at which the airflow is controlled to be generated may be increased. Alternatively, or additionally, an airflow cooling function (not shown) of the appliance 100 may be activated or otherwise controlled to reduce a temperature of the generated airflow. In either case, this may increase the cooling action at the human, and hence reduce the temperature of the user. As another example, if it is determined that the human is relatively cold (for example, the determined temperature is below a second threshold, lower than the first threshold), the speed at which the airflow is controlled to be generated may be decreased. Alternatively, or additionally, an airflow heating function such as a heater (not shown) of the appliance 100 may be activated or otherwise controlled to increase the temperature of the generated airflow. This may act to warm the user and hence increase the temperature of the user. In examples, the control unit 232 may be configured to control the airflow generator 234 (and / or an airflow cooling function and / or an airflow heating function) so that the higher the determined temperature of the human, the faster and / or cooler the generated airflow; and the lower the determined temperature of the human, the slower and / or warmer the generated airflow. The first and second thresholds, the temperature-to-airflow speed mapping, and / or the temperature-to-airflow cooling / warming mapping, may be pre-set or may be set according to the preference of the individual user. Referring to Figure 11, there is illustrated a processing system 1100 according to an example. The processing system 1100 may be used to implement the control unit 232 according to any of the examples described above with reference to Figures 1 to 10. The processing system 1100 comprises a physical processor 1102 and a physical memory 1104. The processing system 1106 comprises an input interface 1106 and an output interface 1108. The processor 1102 may be configured to perform the functions or processes of the control unit 232 according to any of the examples described above with reference to Figures 1 to 10. The memory 1104 may store instructions which, when executed on the processor 1102, cause the processor 1102 to perform the functions or processes of the control unit 232 according to any of the examples described above with reference to Figures 1 to 10. The input interface 1106 may allow the processor 1102 to obtain data, such obtain data representing a thermal image captured by the thermal imaging device 104. The output interface 1108 may allow the processor 1102 to send control signals to other components of the appliance 100. For example, the processor 1102 may control the first unit 102 of the appliance via control commands output via the output interface 1108. In the above examples, a specific appliance 100 is described that has a specific arrangement and form of components. However, it will be appreciated that this need not necessarily be the case, and that in other examples, other appliances having a first unit configured to direct and airflow may be used. In the above examples, the thermal imaging device 104 is attached to the first unit 102. However, it will be appreciated that this need not necessarily be the case and that in other examples the thermal imaging device 104 may be located elsewhere on the appliance 100 or may be provided as separate and / or free-standing component of the appliance 100. In the above examples, a trained ML model is used to determine whether the thermal image represents a human. However, it will be appreciated that this need not necessarily be the case, and that other, non-machine-leaming methods may in principle be used in order to determine whether the thermal image represents a human. For example, the control unit may be configured to analyse the data representing the thermal image and apply a set of rules, for example hard-coded rules, to the data determine whether or not a human is represented in the thermal image. For example, the rules may include that the data represents an object that has a temperature in a range of expected human body temperatures, is of a size within range of expected human sizes, presents with an aspect ratio within a certain range of aspect ratios expected for humans, and / or presents with a shape within a certain range of shapes expected for humans. In the above examples, the trained ML model comprises a human presence classifier that classifies a thermal image according to whether a human is present in the image. However, it will be appreciated that this need not necessarily be the case, and that in other examples, other trained ML models may be used. For example, in other examples, a trained object detection model may be used to simultaneously detect the presence of objects in the image and classify each object according to whether or not the object is a human. In this case, the determination of whether at least a portion of the thermal image represents a human may comprise determining whether a region of the thermal image determined by an object detection model as representing an object, represents a human. Further, it will be appreciated that in some examples any trained ML model may be used to determine whether the thermal image represents a human. For example, the trained ML model may comprise a trained neural network or support vector machine or any other suitably trained ML model. In the above examples, the first unit 102 direct an airflow by rotating the first part 114 relative to the second part 116 about a rotational axis 122. However, it will be appreciated that this need not necessarily be the case, and that other first units may be controlled to direct an airflow to a human in other ways. For example, in other examples, the first unit may be capable of rotation about a plurality of axes for example so as to direct airflow in any direction. Alternatively, or additionally, the first unit may be capable of translation in order to direct the airflow towards a user. Other examples and configurations are possible. Whilst particular examples have thus far been described, it should be understood that these are illustrative only and that various modifications may be made without departing from the scope of the invention as defined by the claims.

Claims

1. An appliance comprising:a first unit configured to direct an airflow;a thermal imaging device; anda control unit;wherein the control unit is configured to:obtain data representing at least a portion of a thermal image captured by the thermal imaging device;determine, based on the obtained data, whether the at least a portion of the thermal image represents a human; andin response to a determination that the at least a portion of the thermal image represents a human, control the first unit to direct the airflow towards the human.

2. The appliance according to claim 1, wherein the control unit is configured to:input the data into a trained machine learning, ML, model, the ML model having been trained to, based on an input of particular thermal image data representing a particular thermal image, output a determination of whether the particular thermal image represents a human; andobtain an output from the trained ML model thereby to determine whether the at least a portion of the thermal image represents a human.

3. The appliance according to claim 2, wherein the ML model comprises a human presence classifier, the human presence classifier having been trained to, based on an input of particular thermal image data representing a particular thermal image, classify the particular thermal image according to whether a human is present in the particular thermal image; and wherein the output comprises the classification.

4. The appliance according to claim 3, wherein the human presence classifier for the ML model has been trained based on training data, the training data comprising a plurality of sets of thermal imaging data each representing a thermal image, each set of thermal imaging data being labelled according to whether the represented thermal image represents a human.

5. The appliance according to any one of claim 1 to claim 5, wherein the control unit is configured to, prior to determining whether the at least a portion of the thermal image represents a human:interpolate the obtained data to increase the resolution of the at least a portion of the thermal image represented by the obtained data.

6. The appliance according to claim 5, wherein the control unit is configured to, prior to determining whether the at least a portion of the thermal image represents a human:apply a smoothing operation to the interpolated data to smooth the at least a portion of the thermal image represented by the obtained data.

7. The appliance according to any one of claim 1 to claim 6, wherein the control unit is configured to:determine, based on the obtained data, a location within the at least a portion of the thermal image at which a human is represented; andcontrol the first unit to direct the airflow towards a location corresponding to the determined location.

8. The appliance according to claim 7, wherein the determination of the location within the at least a portion of the thermal image is responsive to a determination that the at least a portion of the thermal image represents a human.

9. The appliance according to claim 7 or claim 8, wherein the control unit is configured to:determine, based on the obtained data, a location of a centroid of a human represented by the thermal image; anddetermine the location of the centroid as the location within the at least a portion of the thermal image at which a human is represented.

10. The appliance according to claim 9, wherein the location of the centroid of the human is determined by detecting the centre of thermal mass of the human.

11. The appliance according to claim 9 or claim 10, wherein the control unit is configured to:apply a threshold to the obtained data to identify a first part of the at least a portion of the thermal image that represents a human; anddetermine the location of the centroid of the human based on the identified first part.

12. The appliance according to any one of claim 1 to claim 11, wherein the control unit is configured to:determine, based on the obtained data, a distance of one or more humans, represented by the at least a portion of the thermal image, from the thermal imaging device; andcontrol the appliance based on the determined distance.

13. The appliance according to claim 12, wherein the control unit is configured to: determine, based on the obtained data, the distance of a particular human from the thermal imaging device based on a size of a representation of the particular human in the thermal image.

14. The appliance according to claim 12 or claim 13, wherein the control unit is configured to:determine the distance of a first human and a second human, represented by the at least a portion of the thermal image, from the thermal imaging device; andcontrol the first unit to direct the airflow to the first human in preference to the second human on the basis of the distance determined for the first human and the second human.

15. The appliance according to any one of claim 12 to claim 14, wherein the appliance comprises an airflow generator configured to generate the airflow, and wherein the control unit is configured to:determine, based on the determined distance, a particular speed at which airflow is to be controlled to be generated; andcontrol the airflow generator to generate an airflow at the particular speed.

16. The appliance according to any one of claim 1 to claim 15, wherein the control unit is configured to:determine, based on the obtained data, a temperature of a human represented by the thermal image data; andcontrol the operation of the appliance based on the determined temperature.

17. The appliance according to any one of claim 1 to claim 16, wherein the first unit comprises a first part from which the airflow is expelled and a second part, the first unit being configured to direct the airflow by orientating the first part relative to a second part, and wherein the first part comprises the thermal imaging device.

18. The appliance according to claim 17, wherein the thermal imaging device is integrated into, or otherwise attached to, the first part, such that the thermal imaging device is arranged to move simultaneously with the first part relative to the second part.

19. The appliance according to claim 17 or claim 18, wherein a viewing direction of the thermal imaging device is aligned with a direction in which the first part is configured to expel the airflow.

20. The appliance according to claim 19, wherein the control unit is configured to: determine, based on the obtained data, a location within the at least a portion of the thermal image at which a human is represented;determine an offset of the determined location from a central region of the at least a portion of the thermal image; andcontrol the first unit to change the orientation of the first part to reduce the offset.

21. The appliance according to claim 19 or claim 20, wherein the control unit is configured to:determine, based on a first set of obtained data representing at least a portion of a first thermal image captured by the thermal imaging device at a first time, a first location within the at least a portion of the first thermal image at which a human is represented;determine, based on a second set of obtained data representing at least a portion of a second thermal image captured by the thermal imaging device at a second time, a second location within the at least a portion of the second thermal image at which a human is represented;determine, based on the first location and the second location, a direction of travel of the human; andcontrol the first unit to rotate the first part relative to the second part according to the determined direction of travel.

522. The appliance according to any one of claim 19 to claim 21, wherein the control unit is configured to:control the first unit to rotate the first part relative to the second part;in response to the determination that the at least a portion of the thermal image 10 represents a human, control the first unit to stop rotating the first part relative to the second part.

Images