Method for detecting the occupancy of a room and Detector capable of being implemented in such a method

Abstract

A method for detecting the occupancy of a room (R) is provided, the method being implemented by a computer system comprising a processing unit (PU) and a plurality of detectors (D) distributed in a zenithal position in the room (R), a position and an attitude of each detector (D) in the room being defined by detector parameters stored in the computer system. The method comprises a step of preparing a plurality of zenithal images (I) of the room (R) by a plurality of image sensors (IS) of the plurality of detectors (D), at least some of the image sensors (IS) having overlapping fields of view.The method also comprises the step of locating persons in areas of the plurality of images (I) defined by pixel parameters in a coordinate system related to this image and a step of transforming the pixel parameters into second parameters defined in a coordinate system related to the room (R). The method also comprises a step of disambiguating the overlapping areas in order to detect the number and position of the persons occupying the room. The invention also relates to a detector (D) which can be used to implement the method. Figure to be published with the abstract: Fig. 1.

Description

Title of the invention: Method for detecting the occupancy of a room and Detector capable of being implemented in such a method FIELD OF THE INVENTION

[0001] The field of the invention is that of technical management of buildings or home automation. More specifically, the invention relates to a presence detector and to a method for detecting the occupancy of a room or a plurality of rooms in a building. TECHNOLOGICAL BACKGROUND OF THE INVENTION

[0002] Large buildings and industrial sites are generally equipped with a large number of equipment, of various types, which enable the building or site to be occupied and functional. This equipment may, for example, include means of heating and ventilating rooms, means of lighting, secure means of access to some of these rooms, incident detectors (fire). It is sometimes planned to connect at least some of this equipment to a computer supervision system in order to have centralized management of the equipment and buildings. This supervision system may be located on the site itself, or remotely from this site, for example to provide a “cloud” supervision solution.In this context, presence detection (also called occupancy detection or absence detection) is a technology that uses sensors placed in the rooms of the supervised building whose role is to detect human presence (or absence) and, sometimes, to provide a count of the number of people present in a room and in the building. It is also possible, in certain cases, to provide the respective positions of these people, when at least one is present. This presence, absence, count and / or position information (generally referred to as room occupancy detection in this application) can be advantageously used by the supervision computer system to control the operation of certain equipment, for example for energy saving or security purposes.

[0003] Document EP3617933A1 thus discloses an example of a room occupancy detector implementing a plurality of passive infrared sensors, each sensor being capable of providing binary information identifying the presence or absence of a person in its measurement field. The signals provided by these sensors are provided to a classifier configured by learning to establish the count of people present in the room.

[0004] When the number of sensors is relatively small, there is a risk of forming white zones or being unable to discriminate the presence of several people in the measurement field of a sensor. When, on the contrary, a relatively larger number of sensors equip the room in a dense manner, it is possible that several sensors detect the presence of the same person. Also, the solution proposed by the aforementioned document remains imprecise. SUBJECT OF THE INVENTION

[0005] An aim of the invention is to propose a method for detecting the occupancy of a room and an occupancy detector capable of being implemented by this method, overcoming at least some of the limitations of the state of the art. More particularly, an aim of the invention is to provide a reliable solution for detecting people present in a room, that is to say a more precise solution than that proposed by the state of the art in the counting of people present and / or in their location. BRIEF DESCRIPTION OF THE INVENTION

[0006] To this end, the invention relates to a method for detecting the occupancy of a room, the method being implemented by a computer system comprising a processing unit and a plurality of detectors distributed in a zenithal position in the room, a position and an attitude of each detector in the room being defined by detector parameters stored in the computer system.

[0007] According to the invention, the method comprises:

[0008] - a step of preparing a plurality of zenith images of the room, the plurality of images being prepared by a plurality of image sensors respectively forming part of the plurality of detectors, at least some of the fields of vision of the image sensors overlapping; - a step of locating people in areas of the images of the plurality of images to establish a first series of areas, each area located in an image being defined by pixel parameters in a reference frame linked to this image; - A step of transforming the pixel parameters of the zones of the first series of zones using the detector parameters, to establish second parameters of these zones, these second parameters being defined in a reference frame linked to the part; - A step of disambiguating the areas of the first set of overlapping areas by processing the second parameters to provide a second set of areas, the second set of areas being representative of the number and position of people occupying the room.

[0009] According to other non-limiting characteristics of the invention, taken alone, or in any technically feasible combination:

[0010] - the pixel parameters of a marked area in an image comprise the coordinates of a reference point of a bounding box of the image and at least one dimension of this bounding box; - the at least one dimension includes the width, height and / or orientation angle of the bounding box; - the pixel parameters of a marked area in an image also include a degree of presumption of the presence of a person in the bounding box; - the identification step is implemented by at least one classifier receiving as input at least one image of the plurality of images and providing as output the pixel parameters of at least one area identified in this image; - the classifier is a neural network of the SSD or YOLO type; - the detection method comprises a plurality of classifiers implemented by a plurality of processing circuits of the plurality of detectors, a classifier implemented by a processing circuit of a given detector receiving as input an image provided by the image sensor of the given detector and providing as output the pixel parameters of at least one zone of this image; - the detectors each comprise a transmission interface connected to a processing unit of the computer system, the processing unit being configured to implement at least the disambiguation step; - the detection method comprises a step of transmitting the pixel parameters of the areas identified by the plurality of detectors to the processing unit; - the detection method further comprises an initialization step consisting of transmitting the detector parameters to the processing unit and the processing unit is configured to carry out the transformation step; - the processing circuit is also configured to implement the transformation step and the detection method comprises transmitting the second parameters to the processing unit; - at least some of the detectors of the plurality of detectors comprise an attitude sensor, the attitude sensor establishing at least some of the detector parameters; - the disambiguation step implements a non-maximum suppression method.

[0011] The invention also relates to a room occupancy detector, comprising:

[0012] - An image sensor; - A processing circuit connected to the image sensor, the processing circuit being configured to identify at least one area likely to contain a person in an image provided by the image sensor, each area identified in the image being defined by pixel parameters in a reference frame linked to this image; - an attitude sensor, the attitude sensor establishing at least part of the sensor parameters for locating the detector in the room; - a transmission interface for connecting the detector to a processing unit of a computer system.

[0013] According to other non-limiting characteristics of this aspect of the invention, taken alone or in any technically feasible combination:

[0014] - the processing circuit is also configured to transform the pixel parameters of the identified area using the sensor parameters, to establish second parameters of the area, these second parameters being defined in a reference frame linked to the part; - the pixel parameters of a marked area in an image include the coordinates and at least one dimension of a bounding box of the image; - the pixel parameters of a marked area in an image also include a degree of presumption of the presence of a person in the bounding box; - the processing circuit implements a classifier, for example a neural network of the SSD or YOLO type. BRIEF DESCRIPTION OF THE FIGURES

[0015] Other characteristics and advantages of the invention will emerge from the detailed description of the invention which follows with reference to the appended figures in which:

[0016] [Fig.l] [Fig.l] represents a computer system making it possible to implement a detection method in accordance with the invention;

[0017] [Fig.2] [Fig.2] illustrates the location of a detector in a frame linked to the world;

[0018] [Fig.3] [Fig.3] schematically represents the architecture of a detector in accordance with one aspect of the invention;

[0019] [Fig.4]

[0020] [Fig.4] schematically represents a classifier which can be implemented by a processing circuit of a detector according to one aspect of the invention;

[0021] [Fig.5] [Fig.5] represents a projection on the ground of the fields of vision of the dice detectors equipping the part shown in [Fig.l];

[0022] [Fig.6a]

[0023] [Fig.6b]

[0024] [Fig.6c]

[0025] [Fig.6d] Figures 6a to 6d represent the successive steps of a detection method according to one aspect of the invention; DETAILED DESCRIPTION OF THE INVENTION General presentation of the computer system

[0026] [Fig. 1] represents a computer system making it possible to implement a method for detecting the occupancy of a room R in accordance with the invention. It is recalled that in the context of the present detailed description, in addition to a simple detection of the presence or absence of a person, such occupancy detection can also correspond to counting the number of people in this room R and / or the location of these people.

[0027] For the sake of simplification, [Fig.l] shows a computer system with which a single room R of a building has been equipped, but the invention applies to any enclosed space (for example a building) whether this space is composed of one room or a plurality of such rooms, or even to the detection of the occupation of an unenclosed space (for example a stadium, a garden). The term "room" therefore very generally designates any space likely to be occupied by people and whose occupancy is to be monitored.

[0028] The computer system is composed of a plurality of detectors D, distributed in zenithal positions in the room R, fixed to the ceiling of the room R (or arranged for example on masts if the room is an open space). As will be detailed in the rest of this description, each detector D is provided with an image sensor whose optical axis of view is oriented towards the floor of the room R, the field of vision of each image sensor therefore intercepting the floor to define an exposed surface of the floor FOV. It is very generally sought that the agglomeration of the exposed surfaces FOV covers the floor of the room R as much as possible, at least in the parts of this room which are likely to be occupied by people, so as to leave no white zone. This therefore often leads to the fields of vision (and therefore the exposed surfaces FOV of the floor) of the image sensors of certain detectors D overlapping, as can be seen in [Fig.l].The exposed FOV surfaces can, depending on the nature of the image sensor (in particular the focal length of its optics) and the height positioning of the detector D, have ground dimensions typically between 1mA2 and 25mA2.

[0029] In a very general manner, the invention aims to exploit, by computer image processing techniques, the images provided by the network of detectors D arranged in the room R according to the arrangement which has just been described, so as to locate people. In this way, by taking care to take into account the overlapping fields of vision of certain D detectors, it is possible to count and / or locate the number of people occupying the room, without necessarily deploying a large number of D detectors.

[0030] For this purpose, the detectors D comprising the image sensors are connected to a processing unit PU of the computer system, which is here arranged "in the clouds". This processing unit PU is provided, in a completely conventional manner, with elementary processing components (CPU, memory, input / output interfaces, etc.) arranged together to implement a wide variety of computer processing operations, in particular those described in the remainder of this description. The computer means of the system (i.e. at least the detectors D and the processing unit PU) collaborate to implement the method for detecting the occupancy of the room R. Other configurations of the computer system can of course be envisaged. A local router LR can thus be provided, as is the case in the configuration shown in [Fig.l], or a plurality of such routers to aggregate the information provided by the detectors D before transmitting it to the processing unit PU. The processing unit PU is not necessarily located at a great distance from the building, and it may in particular reside in the room R itself, or in another room of the building.

[0031] The computer system makes it possible in particular to synchronize the time of acquisition of the images by the detectors D, so that these are representative of the state of occupancy of the room R at a given time. This may involve configuring or communicating a clock common to all the detectors D, an instant or instants of image acquisition being planned in advance relative to this clock, i.e. at a fixed time. Alternatively, it may involve simultaneously addressing to the detectors D an event triggering acquisition of an image. The acquisition of images may be repeated over time, for example at a frequency of between 1 time per second and 1 time per hour, in order to continuously establish the information on occupancy of the room or the count of people occupying it.Regardless of how the image capture is synchronized and at what rate this image capture is repeated, these aspects can be implemented by a computing routine running on the processing unit PU, on a local router LR or any other computing element composing the computing system.

[0032] With reference to Figures 1 and 2, a detector D equipping the part R is characterized by its position and its attitude in a reference frame linked to this part (0, x, y, z), sometimes designated by the expression “world reference frame”. Thus, the position PD of a detector D is determined by its coordinates (xD, yD, zD) in this reference frame. This is typically the position of the detector D according to the length, the width and the height in the part R, when the reference frame linked to the part R is aligned along the sides of this part as is the case in [Fig.l]. More generally, the Oz axis of the reference frame linked to the part is vertical, and the Ox, Oy axes are horizontal and perpendicular to each other. The attitude AD of the detector D in this reference frame is determined by angles called roll rD, pitch tD and yaw 1D, these angles being respectively defined by the angular orientation around the Ox, Oy and Oz axes of a reference frame (0, x', y', z') linked to the detector D. Preferably, this reference frame is positioned to reside in an image plane of the image sensor, and will be referred to as the "image-linked reference frame" in the remainder of this description. Furthermore, the expression "detector parameter" will designate at least the position Pd (xd, yD, zD) and the attitude AD(rD, tD, 1D) of a detector D in the reference frame linked to the part, as well as, preferably, the other "intrinsic" parameters of the image sensor.

[0033] The occupancy detection processing implemented by the computer system requires calibrating the image sensors of the detector network D, i.e. determining the point transformation existing between a pixel of an image prepared by the image sensor of a detector D and a point in the reference frame linked to the part. This transformation comprises the application of a translation in position and a rotation in attitude aimed at matching the reference frame linked to the image and the reference frame linked to the part. The translation and rotation to be applied to calibrate a detector D are given by the parameters PD, AD of this detector D.

[0034] As is well known per se, other parameters may be useful to precisely define the transformation of a point defined in the part coordinate system and a pixel. These additional parameters include the so-called "extrinsic" parameters of the image sensor (focal length, optical center and tilt coefficients) which define the projection of the part coordinate system into the image coordinate system (pixel matrix). These other parameters also include the lens distortion parameters, which are used to compensate for the distortion introduced by the lens associated with the image sensor.

[0035] For this calibration purpose, the position parameters PD and attitude AD of each detector D (and more generally all the calibration parameters) are recorded in the computer system, for example in a file, or a database, retained on the processing unit PU or a local router LR. Alternatively, these parameters can be respectively recorded in the detectors D.

[0036] Some of these parameters can be determined visually when installing the detectors in the room R. This concerns in particular the position PD of the detector D and its yaw angle 1D which can be easily measured, to within a few centimeters or a few degrees without this imprecision affecting the results of the detection. As for the 1D yaw angle, the installation of a D detector in a room may require orienting a visual orientation marker RO placed on the D detectors (visible in [Fig.2]) in a specific direction, so that the explicit measurement of this angle is not necessary, but considered as an imposed datum.

[0037] Other parameters of the detectors D must, however, be determined with more precision than by simple visual inspection or measurement, since the detection result can be particularly sensitive to the precision of their values. These include the roll angle rD and pitch angle tD, which determine the direction of the viewing angle of the image sensor. These angles affect the field of vision of the image sensor all the more the further the detector is from the floor of the room. To address this problem, a detector D according to the invention is provided with an attitude sensor capable of determining with sufficient precision, preferably less than 1° or less than 2°, these roll and pitch angles, as detailed in more detail in the following section of this description.

[0038] For the extrinsic and distortion parameters of the lens, it is possible to use predetermined values, for example default values ​​provided by the manufacturer. However, it is preferable to determine these values ​​individually for each detector D with greater precision before installing the detectors in the room, and to store them in the computer system as mentioned above. For the purposes of this description, these parameters will be considered available and used in all transformations that will be applied to the images. Detector

[0039] [Fig. 3] shows the internal architecture of a detector D according to one aspect of the invention. As already stated, it comprises an image sensor IS capable of forming an image of the environment arranged in its field of vision, that is to say the portion of the space perceived by this sensor IS when its optical axis is oriented in a determined direction. As is well known per se, an image established by such a sensor is formed of a plurality of pixels, arranged in rows and columns. A point of this image can therefore be identified by rows, according to a row and a column of pixels, in the reference frame (O', x', y') linked to this image.More generally, a portion of an image (also referred to as an "area" in this application) may be defined by pixel parameters in this reference system, for example the coordinates of the center of the area, and a principal dimension (in pixels) of the extent of the area, for example a diameter or a side, depending on the shape of the area.

[0040] The IS image sensor can operate in a wide variety of spectral bands but, advantageously and to facilitate the detection of people in the image, the IS image sensor operates in the infrared. Each pixel can be defined by a single value (gray level or intensity level) or by a plurality of values ​​(e.g. RGB values).

[0041] The detector D of [Fig. 3] also comprises, functionally connected to the image sensor IS, a processing circuit PC. The images captured by the image sensor IS are supplied to the processing circuit PC, and the latter is configured to identify in the image at least one area likely to contain a person. This configuration can be hardware and / or software, and in the latter case a program executes on elementary circuits (microcontroller, memory, etc.) of which the processing circuit PC is composed.

[0042] The location in the image produced by the processing circuit PC can be implemented by a classifier K, for example a neural network of the SSD or YOLO type, configured by training from reference images. Such a classifier, as shown schematically in [Fig. 4], receives as input (typically via convolution layers) the image I prepared by an image sensor IS, and provides as output groups of parameters BB; (i=1 to n), each group defining an area of ​​the image I likely to contain a person. Reference may be made to the abundant literature in the field to produce such a classifier, for example to the article “SSD: single shot multibox detector”, Wei Liu et al, Computer Vision - ECCV 2016. ECCV 2016. Lecture Notes in Computer Science, vol 9905. Springer, Cham.

[0043] It is usual to define an area identified by these methods in the form of a bounding box in the image, such a box being defined by the coordinates c; (in the matrix of pixels forming the image) of the center of the box, at least one dimension (side or width w; and height h; of the area, expressed in pixels) of this box. Other representations are possible. It is also possible, for greater precision, to provide additional parameters, such as the orientation of the box in the image, for example the angle formed by one side of the area with respect to a reference direction in a reference frame linked to the image I.

[0044] In all cases, and regardless of the manner in which a bounding box is identified in the image I by the classifier K, such a box is defined by parameters BB; called “pixels” in a reference frame linked to the image I.

[0045] When the image registration is implemented by a classifier K, the latter typically has a plurality of output groups. Each group of outputs provides the pixel parameters BB; of a zone registered in the image, on respective outputs of the group, for example the coordinates of the center Ci, a width Wi and a height h; of a bounding box, as shown in [Fig.4] for the n groups of parameters BBi,...,BBn. The pixel parameters of a group of pa BB meters; may also include a degree a; of presumption of the presence of a person in the identified area. Typically, 20 to 40 groups of outputs, or even more, can be provided so that the K classifier can identify 20 to 40 areas in an image likely to contain a person. Only areas are retained for which the degree a; of presumption of the presence of a person is greater than a determined threshold, for example greater than 0.5 when this degree is evaluated between 0 and 1.

[0046] In any event, the processing circuit PC of a detector D according to the invention is capable of establishing zones of the image provided by the image sensor IS, these zones being defined by pixel parameters in a reference frame linked to the image, each zone being capable of containing a person. The processing circuit PC may have other functions, such as that aimed at coordinating the different elements making up the detector, or that aimed at controlling the image sensor, for example to trigger a shot. Alternatively, these additional functions may be implemented by a dedicated microcontroller of the detector D, connected to the different elements making it up.

[0047] Continuing the description of [Fig. 3], the processing circuit PC is functionally connected to a transmission interface PO which makes it possible to connect the detector D directly or indirectly to the calculation unit PU of the computer system. Thus, information collected individually by the detectors D, in particular the parameters BB; areas identified in an image, can be transmitted to this processing unit PU. Corolarily, the processing unit PU can transmit information to each detector D of the computer system, for example to configure them. This networking of the detectors D and the processing unit PU can be implemented according to any suitable protocol. The connection between the detectors D and the processing unit PU can be wired or wireless.For example, the detectors D may be individually wired to a local router LR, this router being capable of aggregating the transmitted information to transmit it to the processing unit PU via a network, for example the Internet network. Rather than providing for the wiring of the detectors D to the local router, it may be preferable, in order to facilitate the implementation of the computer system, to connect these detectors D to the router by a short-distance wireless link such as a Bluetooth™ link. Alternatively, the transmission interface FO of the detectors D is capable of operating a long-distance wireless transmission and therefore of transmitting the collected data to the processing unit PU without the need for a local router. Such a long-distance wireless link may, for example, implement LORA, SIGFOX or 5G technology.

[0048] According to an important characteristic of a detector D according to the invention, and as was briefly presented in the previous section, this comprises an attitude sensor OS, that is to say a sensor capable of locating the orientation of the detector in space (and therefore of the image sensor IS), in a frame of reference linked to the part R. This may be a micro-electromechanical device (MEMS according to the English acronym) capable of providing static attitude information (roll, pitch and yaw angles) typically with an accuracy of less than 2° or 1°, or even less than 0.5°. The attitude sensor OS is at least connected to the transmission interface OS, so that the attitude parameters AD provided by this sensor OS can be communicated to the processing unit PU of the computer system to be stored there. This communication can be carried out upon initialization of the detector D, for example during its startup sequence, or upon express request from the processing unit PU.

[0049] Of course, the detectors D are provided with elements ensuring their electrical power supplies, for example connectors allowing their connection to a power supply network and / or an energy storage element. Method for detecting room occupancy

[0050] The detection method takes advantage of the computer system which has just been described, this system comprising the processing unit PU and the plurality of detectors D distributed in a zenithal position in the room. At least some of the fields of vision of the image sensors IS of the detectors overlap, as has been shown in [Fig.l]. Thus [Fig.5] shows the projection onto the ground of the fields of vision of the detectors D equipping the room R shown in [Fig.l], these projections forming 5 exposed surfaces FOV1-FOV5, in correspondence with the 5 detectors D of [Fig.l]. In this [Fig.5], 3 people PI, P2, P3 occupying the room R are symbolized by dotted circles.

[0051] During the physical installation of the detectors D in the room R, their respective positions PD were identified. The detectors D were also positioned in the room R by orienting them according to a determined 1D yaw angle, using for example the visual reference RO with which they may be equipped. These position and yaw orientation data, which form part of the parameters of the detectors D, can be stored in the computer system. During a preliminary initialization phase of the computer system, the detectors D can also broadcast their attitude parameters, and in particular their pitch tD and roll rD, to the computer system, so that all the parameters of the detectors D (position and attitude) are stored and available in the computer system, for example at the processing unit PU.Other approaches to collect and store the detector parameter D in the computer system are possible, for example the 1D yaw angle provided by the attitude sensor AS can be used instead of considering this angle. as a fixed and predefined value.

[0052] The detection method comprises a first step of preparing a plurality of zenith images of the room, the plurality of images being prepared by the plurality of image sensors IS respectively forming part of the plurality of detectors D. These images are produced at a determined instant, this instant being coordinated by means of a common clock or triggered by an event broadcast in the computer system, as previously specified.

[0053] [Fig.6a] represents the 5 images II to 15 acquired by the 5 image sensors IS of the 5 detectors D with which the room R of FIGS. 1 and 5 was equipped. These images, in particular when they come from infrared image sensors IS, have an increased intensity in areas 111, 112, 121, 141, 151 corresponding to the location of a person. It is noted that, in [Fig.6c], people PI, P3 located in the overlap of the fields of vision of a plurality of sensors, there are more image areas with an increased intensity (5 in the present case) than people actually in the room R.

[0054] The method comprises, following the first preparation step, a step of locating people in areas of the images. This step, as illustrated in [Fig.6b], leads to determining in each of the images 11-15 the areas, represented in this figure in the form of dotted bounding boxes, likely to contain people. In [Fig.5], we find 5 bounding boxes respectively defined by their pixel parameters BB11, BB12, BB21, BB41, BB51. The boxes are respectively positioned on the areas 111, 112, 121, 141, 151 corresponding to the location of a person in [Fig.6a].

[0055] When the detectors D are in accordance with that presented in the previous section of this description, this identification step can be implemented by the processing circuit, and in particular by the classifier K implemented by this circuit, and to which the image has been communicated. In this case, each detector D processes, via the classifier K of its processing circuit PC, the image provided by its image detector IS. All of the images are therefore processed during this identification step in a distributed manner over all of the detectors D.

[0056] Alternatively, it may be provided that this tracking step is implemented by the processing unit PU of the computer system, which requires an intermediate step of transferring images to this unit PU, via the I / O transmission interfaces of the detectors D. This approach is however not the preferred approach insofar as it requires the transfer of a relatively large quantity of information to this unit PU, but it nevertheless remains possible.

[0057] Whether the tracking processing is carried out by the detectors in a distributed manner by the detectors D or centralized by the processing unit PU, we have at the end of this step of a first series of zones BB11-BB51 likely to contain people P1-P3, each zone identified in an image being defined by pixel parameters in a reference frame linked to this image.

[0058] As indicated previously, these zones are preferably bounding boxes and the pixel parameters of such a zone preferably comprise the coordinates of a reference point (for example the pixel coordinates of the center) of the bounding box and at least one pixel dimension of this box (a side if this box is square, a length and a width if this box is rectangular). It can also be provided that the parameters comprise an orientation angle of the bounding box and / or also a degree of presumption of the presence of a person in this box (typically a value between 0 and 1). It is noted that at this stage the same person, for example the person referenced P2 in [Fig. 5], can be located in two zones of two distinct images, the reference bounding boxes BB12 and BB21 in [Fig. 6b].

[0059] In a following step called "transformation" the pixel parameters of the zones of the first series of zones are transformed using the parameters of the detectors D, to establish second parameters of these zones, the second parameters being defined in a reference frame linked to the part. In other words, the detector parameters D (position and attitude of this detector in the reference frame linked to the part R, but also the other parameters, including the extrinsic parameters and the distortion parameters) are used to transform the pixel parameters of the zones (in the reference frame specific to each image therefore) into parameters which are expressed in the reference frame linked to the part R.

[0060] Preferably, this transformation is carried out by the processing unit PU and therefore, between the identification step and the transformation step, a step of transmitting the pixel parameters of the zones identified by the plurality of detectors D to the processing unit PU has been provided. It is noted that this transmission requires a much lower bandwidth than that which would have been necessary for the transmission of the complete images. This therefore makes it possible to limit the use of the network of the computer system and therefore, at constant capacity, to increase the number of detectors D in this system or the frequency of shooting.

[0061] Alternatively, it may be provided that this transformation is carried out by the processing circuit PC of each detector. In this case, provision will have been made, for example during the preliminary step of initializing the computer system, for the transmission to the processing circuits PC of each detector of its sensor parameters, i.e. at least the position of these detectors D in a reference frame linked to the room, which the detector D can combine with the attitude parameters provided by the attitude sensor OS to carry out this transformation. The second parameters describing, in a coordinate system linked to the room, the areas likely to contain people are then transmitted to the PC processing unit by each of the I / O transmission circuits of the D detectors.

[0062] At the end of this step, and whatever the approach chosen, the processing unit PU has the parameters of the zones likely to contain people, these parameters being expressed in the same reference frame linked to the room, which makes it possible to represent them in the same reference frame as shown in [Fig.6c]. As can be seen in this figure, among the 5 zones of the first series of zones, the two zones referenced B' 12 and BB'21 and the two zones referenced BB'41, BB'51 correspond respectively substantially to the same regions of the room R.

[0063] Finally, following the transformation step, the method comprises a step of disambiguating the zones of the first series of overlapping zones. This step may in particular implement a non-maximum suppression method, well known in the field of computer vision, and a detailed description of which can be found in the document “Non-Maximum Suppression for Object Detection by Passing Messages between Windows”, April 2015, Asian Conference on Computer Vision. This step aims to eliminate duplicated zones, i.e. zones present in the overlapping parts of two distinct images, and in fact corresponding to the same person. This step is carried out by processing the second parameters of the zones, in the reference frame linked to the part, to provide a second series of zones not presenting “duplicate” zones.This second series of zones therefore corresponds to the first series from which the redundant zones corresponding to the same person present in the overlapping fields of vision of two detectors D have been filtered. The second series of zones established during this disambiguation step is therefore representative of the position of the people occupying the room, and the number of zones in this second series therefore corresponds to the number of people present in the room. Thus, in [Fig.6d], we have represented the 3 zones, here the bounding boxes BB'1, BB'2, BB'3, composing the second series of zones after application of the disambiguation step to the first series of zones represented in [Fig.6c]. We find three bounding boxes BB'1, BB'2, BB'3 respectively arranged in correspondence (and in the same reference frame) of the 3 people represented in [Fig.5]. It is therefore possible to detect the presence of these 3 people, to count them and to locate them.

[0064] As already specified, the steps of preparation, identification, transformation and disambiguation can be repeated cyclically, in order to keep the room occupancy information up to date.

[0065] It is noted that this disambiguation step, to be effective, requires that identified areas from different images overlap (corresponding to the same person located in the room R) are superimposed on each other with good precision after transformation of their pixel parameters in the reference frame linked to the room. This good precision is made possible, according to the invention, by the precise determination of the attitude of the detectors (and more precisely of the image sensors of these detectors) in the room by means of the attitude sensor AS. Such precision could not be easily obtained by a simple visual measurement of this attitude.

[0066] It is further noted that the detection method does not require transferring images into the network of the computer system, nor does it require re-aligning the images with each other to take into account the overlapping of some of these images with each other, which could be computationally cumbersome to implement.

[0067] Of course, the invention is not limited to the embodiments described and variant embodiments can be made without departing from the scope of the invention as defined by the claims.

Claims

Claims

1. A method for detecting the occupancy of a room (R), the method being implemented by a computer system comprising a processing unit (PU) and a plurality of detectors (D) distributed in a zenithal position in the room (R), a position and an attitude of each detector (D) in the room being defined by detector parameters stored in the computer system, the method comprising: - a step of preparing a plurality of zenithal images (I) of the room (R), the plurality of images (I) being prepared by a plurality of image sensors (IS) respectively forming part of the plurality of detectors (D), at least some of the fields of vision of the image sensors (IS) overlapping;- a step of locating people in areas of the images of the plurality of images (I) to establish a first series of areas, each area located in an image (I) being defined by pixel parameters in a reference frame linked to this image; - A step of transforming the pixel parameters of the areas of the first series of areas using the detector parameters, to establish second parameters of these areas, these second parameters being defined in a reference frame linked to the room (R); - A step of disambiguating the areas of the first series of overlapping areas by processing the second parameters to provide a second series of areas, the second series of areas being representative of the number and position of the people occupying the room.;

2. Detection method according to the preceding claim in which the pixel parameters of a zone identified in an image (I) comprise the coordinates of a reference point of a bounding box (BB) of the image (I) and at least one dimension of this bounding box (BB).

3. Detection method according to the preceding claim, in which the at least one dimension comprises the width, the height and / or the orientation angle of the bounding box (BB).

4. Detection method according to one of the two preceding claims in which the pixel parameters of an area identified in an image (I) also include a degree of presumption of the presence of a person in the bounding box.

5. Detection method according to one of the preceding claims in which the identification step is implemented by at least one classifier (K) receiving as input at least one image (I) of the plurality of images and providing as output the pixel parameters of at least one area identified in this image.

6. Detection method according to the preceding claim comprising a plurality of classifiers (K) implemented by a plurality of processing circuits (PC) of the plurality of detectors (D), a classifier (K) implemented by a processing circuit (PC) of a given detector (D) receiving as input an image (I) provided by the image sensor (IS) of the given detector and providing as output the pixel parameters of at least one zone of this image (I).

7. Detection method according to the preceding claim in which the detectors (D) each comprise a transmission interface (I / O) connected to a processing unit (PU) of the computer system, the processing unit (PU) being configured to implement at least the disambiguation step.

8. Detection method according to the preceding claim comprising a step of transmitting the pixel parameters of the areas identified by the plurality of detectors (D) to the processing unit (PU).

9. A detection method according to the preceding claim further comprising an initialization step of transmitting the detector parameters to the processing unit (PU) and wherein the processing unit (PU) is configured to perform the transformation step.

10. Detection method according to one of claims 6 or 7 in which the processing circuit (PC) is also configured to implement the transformation step and the detection method comprises the transmission of the second parameters to the processing unit (PU).

11. A detection method according to one of the preceding claims wherein at least some of the detectors (D) of the plurality of detectors comprise an attitude sensor (OS), the attitude sensor (OS) establishing at least part of the detector parameters.

12. Room occupancy detector (D), comprising: - An image sensor (IS); - A processing circuit (PC) connected to the image sensor (IS), the processing circuit (PC) being configured to identify at least one area likely to contain a person in an image (I) provided by the image sensor (IS), each area identified in the image (I) being defined by pixel parameters in a reference frame linked to this image (I); - an attitude sensor (OS), the attitude sensor (OS) establishing at least part of the sensor parameters making it possible to locate the detector (D) in the room (R); - a transmission interface (I / O) for connecting the detector (D) to a processing unit (PU) of a computer system.

13. Occupancy detector (D) according to the preceding claim, in which the processing circuit (PC) is also configured to transform the pixel parameters of the identified area using the sensor parameters, to establish second parameters of the area, these second parameters being defined in a reference frame linked to the room (R).

14. Occupancy detector (D) according to one of claims 12 to 13 in which the pixel parameters of a zone identified in an image (I) comprise the coordinates and at least one dimension of a bounding box (BB) of the image (I).

15. Occupancy detector (D) according to one of claims 12 to 14 in which the processing circuit implements a classifier (K), for example a neural network of the SSD or YOLO type.

Images