Method and system for anonymizing images from a video surveillance camera

Abstract

An anonymization system for at least one raw image (IMb) of a scene including at least one individual, the anonymization system comprising: at least one CCTV camera configured to acquire a raw image (IMb) representative of the scene; at least one detection device configured to determine a point cloud (NP) representative of the scene;a computer configured to: identify at least one sub-image (IMs) representative of an individual (9) in the raw image (IMb), the sub-image (IMs) having a primary position (P1) in the scene (S), identify at least one sub-point cloud (NPs) representative of an individual (9) in the point cloud (NP), the sub-point cloud (NPs) having a secondary position (P2) in the scene, and anonymize the sub-image (IMs) of the raw image (IMb) so as to obtain an anonymized image (IMa) if a gap (EP), determined between the primary position (P1) and the secondary position (P2), is less than a predetermined threshold (S1). Abstract figure: Figure 7;

Description

Title of the invention: Method and system for anonymizing images from a video surveillance camera technical field

[0001] The present invention relates to the field of video surveillance and video protection of a scene in which an individual is moving, particularly in a railway environment, for example in a railway station, at a level crossing or directly inside a railway vehicle. More particularly, the present invention relates to the anonymization of images from a video surveillance camera.

[0002] In order to ensure the safety of users and / or obtain statistical information to improve services, it is known to position surveillance cameras on or in a railway vehicle or in a railway station. Such surveillance cameras are generally RGB type cameras, meaning Red Green Blue, and designating a camera capable of capturing color images.

[0003] The use of images from surveillance cameras is restrictive because they may infringe on the privacy of individuals identifiable in the images. Without the express consent of these individuals, the images must be processed by sworn officers solely for security purposes.

[0004] Also, it is known to proceed with anonymization of images from the video surveillance camera in order to ensure that no person is identifiable.

[0005] For this purpose, manual anonymization is known, carried out image by image by a sworn agent.

[0006] However, when dealing with a large number of images to be anonymized, manual processing is time-consuming and requires the presence of several sworn officers. In particular, an autonomous train includes a very large number of CCTV cameras to ensure optimal operation, which generates a quantity of data to be processed that is too large to be handled manually.

[0007] Also known is a so-called "semi-automatic" anonymization, performed by computer using specific algorithms. However, it is necessary for the computer-processed images to be checked, which similarly requires the presence of several sworn officers.

[0008] Moreover, in the event of poor visibility, due for example to low light, some data may not be detected by the specific algorithms, which leads to processing defects in images from video surveillance cameras. Detection failures can occur particularly in darkness, for example at night.

[0009] The invention thus aims to eliminate at least some of these drawbacks by proposing a system and a method for anonymizing images from at least one effective and reliable video surveillance camera, which makes it possible to ensure the detection of an individual even in the event of low visibility. PRESENTATION OF THE INVENTION

[0010] The invention relates to a system for anonymizing at least one raw image of a scene comprising at least one individual, the anonymization system comprising: • at least one CCTV camera configured to acquire a representative raw image of the scene, • at least one detection device configured to determine a point cloud representative of the scene, • a calculator configured for: • identify at least one representative sub-image of an individual within the raw image, the sub-image having a primary position in the scene, • identify at least one sub-cloud of points representative of an individual in the point cloud, the sub-cloud of points having a second position in the scene, • anonymize the sub-image of the raw image so as to obtain an anonymized image if a gap, determined between the first position and the second position, is less than a predetermined threshold.

[0011] The anonymization system according to the invention advantageously makes it possible to ensure the detection of an individual in a scene, even under poor conditions, for example in darkness, by correlating the data acquired by the video surveillance camera and that acquired by the detection device. Thus, a partial detection by one of the video surveillance cameras or the detection device can be confirmed by the other of the video surveillance camera or the detection device, allowing for reliable detection.

[0012] According to a preferred aspect, the detection device is a remote sensing sensor of the LIDAR, radar, infrared type, allowing the use of a device integrating metrological data and capable of detecting an object or an individual regardless of visibility or brightness.

[0013] Alternatively, the detection device comprises at least two video cameras configured to determine a three-dimensional point cloud from of two images captured from two different positions. Such a three-dimensional determination is achieved by stereovision or by the principle known as Structure From Motion, meaning "Structure acquired from a movement" in English and corresponding to a principle of photogrammetric interval imaging.

[0014] In one embodiment, the anonymization system comprises a plurality of CCTV cameras, each CCTV camera being configured to acquire a raw image of the scene, thereby increasing the reliability of individual detection. When the raw images are acquired from different angles, this helps to ensure the detection of an individual present.

[0015] In an alternative or complementary embodiment, the anonymization system comprises a plurality of detection devices, each detection device being configured to determine a point cloud of the scene, thereby increasing the reliability of individual detection. When point clouds are acquired from different angles, this ensures the detection of an individual present.

[0016] The invention further relates to a railway vehicle comprising at least one anonymization system as described above.

[0017] The invention also relates to a method for anonymizing at least one raw image of a scene comprising at least one individual, the method being implemented by an anonymization system as described above, the method comprising the steps of: • identify at least one representative sub-image of an individual in the raw image acquired by a CCTV camera, the sub-image having a primary position in the scene, • identify at least one sub-cloud of points representative of an individual within a point cloud determined by a detection device, the sub-cloud of points having a secondary position in the scene, • determine a difference between the primary position and the secondary position, and • anonymize the sub-image of the raw image so as to obtain an anonymized image if the difference is less than a predetermined threshold.

[0018] The anonymization process advantageously uses, in addition to a video surveillance camera, a detection device that allows the addition of metrological information that characterizes visual information. Thus, even in the event of reduced visibility of the video surveillance camera, due for example to poor weather conditions or low light, the detection device makes it possible to detect the presence of an individual.

[0019] In one embodiment, the detection device is a remote sensing sensor of the LIDAR, radar, infrared type, the method thus uses an "active sensor", which allows the addition of metrological information in a simple and direct manner.

[0020] Thus, the anonymization process according to the invention makes it possible to ensure the detection of an individual by using two different sensors and correlating the information captured by each.

[0021] More reliable detection can advantageously be automated, thereby reducing the need for a sworn officer and allowing for the processing of a larger number of images. This is particularly advantageous for an autonomous train, which requires a very large number of CCTV cameras to operate.

[0022] In one embodiment, the method comprises the steps of: • identify an initial colorimetry and / or reflectivity of the sub-image, • identify a second colorimetry and / or a second reflectivity of the sub-point cloud, and • determine a difference between the first colorimetry and / or the first reflectivity and the second colorimetry and / or the second reflectivity.

[0023] This increases the reliability of comparison between the sub-image identified in the raw image and the sub-point cloud identified in the point cloud, thereby increasing the reliability of detecting an individual.

[0024] In one embodiment, the method comprises the steps of: • identify a first shape in the sub-image, • identify a second shape in the subcloud of points, and • determine a difference between the first shape and the second shape.

[0025] This makes it possible to increase the accuracy of detecting one or more individuals. The detection is thus more robust.

[0026] The invention also relates to a computer program type product, comprising at least one sequence of instructions stored and readable by a computer and which, once read by this computer, causes the execution of the steps of the process as described above.

[0027] The invention further relates to a computer-readable medium comprising the computer program-type product as described above. PRESENTATION OF THE FIGURES

[0028] The invention will be better understood upon reading the following description, given by way of example, and referring to the following figures, given by way of non-limiting examples, in which identical references are given to similar objects.

[0029] Fig. 1 is a schematic representation of a railway vehicle comprising an anonymization system according to one embodiment of the invention.

[0030] Fig. 2 is a schematic representation of an anonymization system according to an alternative embodiment of the invention.

[0031] The [Fig.3] is a schematic representation of a raw image of a scene acquired by a video surveillance camera of the anonymization system of the [Fig.2].

[0032] The [Fig.4] is a schematic representation of a point cloud determined by a detection device of the anonymization system of the [Fig.2].

[0033] Fig. 5 is a schematic representation of the point cloud of Fig. 4 projected onto the raw image of Fig. 3.

[0034] [Fig.6] is a schematic representation of an anonymized image of the scene in [Fig.3].

[0035] Fig. 7 is a diagram of the steps of an anonymization process according to one implementation method of the invention.

[0036] Fig. 8 is a diagram of an anonymization process in the case of a plurality of video surveillance cameras.

[0037] It should be noted that the figures set out the invention in detail to implement the invention, said figures being of course able to serve to better define the invention where appropriate. DETAILED DESCRIPTION OF THE INVENTION

[0038] The invention relates to the anonymization of images from a video surveillance camera, particularly in the railway sector. The invention is presented in a railway context but it is applicable to any place open to the public, in particular, an airport, a bus station, a port, a public building, etc.

[0039] With reference to [Fig. 1], a railway vehicle 1 is shown comprising an anonymization system 2 according to one embodiment of the invention. In this example, the anonymization system 2 is mounted on the railway vehicle 1 to view a scene S in the vicinity of the railway vehicle 1, for example, in a station or on the railway tracks, in order to secure the environment of the railway vehicle 1. By way of example, the anonymization system 2 makes it possible to capture a scene S near a level crossing. It is understood that the anonymization system 2 could alternatively be mounted inside the railway vehicle 1.

[0040] The anonymization system 2 according to the invention is of particular interest for securing the environment of autonomous trains.

[0041] For the sake of clarity in the figures, the anonymization system 2 will be described hereafter according to a second embodiment of the invention, represented in [Fig.2], in which it is mounted in a station for example, to allow visualization of a scene S near the railway vehicle 1.

[0042] With reference to Figures 1 and 2, the anonymization system 2 comprises a video surveillance camera 3, a detection device 4 capable of generating a three-dimensional representation and a computer 5, connected by a data network to the video surveillance camera 3 and to the detection device 4.

[0043] The video surveillance camera 3 is configured to view a scene S potentially containing one or more individuals 9 and to acquire a raw image IMb of the scene S, illustrated in [Fig. 3]. In this example, the raw image IMb contains an identifiable individual 9.

[0044] In practice, the raw IMb image is defined in two dimensions represented, in this example, in an orthogonal plane (X, Y), as illustrated in [Fig.3].

[0045] In this example, the video surveillance camera 3 is an RGB camera, meaning "Red Green Blue" in English, and designating a camera that uses the three primary colors to transmit and reconstruct the raw image IMb. It goes without saying that the video surveillance camera 3 could be different. Similarly, the term "video surveillance camera" refers to any image acquisition device.

[0046] The detection device 4 is configured to determine a three-dimensional NP point cloud representative of the scene S and illustrated in [Fig. 4]. To this end, the detection device 4 is preferably configured to emit radiation into the scene S, receive a response signal from the emitted radiation, and generate an NP point cloud representative of the received signal in response to the emitted radiation.

[0047] In other words, the detection device 4 is preferably a sensor capable of integrating metrological data of the scene S. This is commonly referred to as an "active" sensor. In this example, the detection device 4 is a LIDAR-type remote sensing sensor, meaning "laser imaging detection and ranging" or "light-based distance estimation." It goes without saying that the detection device 4 could be different, for example, a radar, infrared, or any other type of sensor integrating metrological data.

[0048] In an alternative embodiment (not shown), the detection device 4 comprises a plurality of video cameras capable of reconstructing the scene in three dimensions from different two-dimensional images acquired from different positions and therefore different angles. Such reconstruction makes it possible to measure distances which allow the generation of the point cloud NP.

[0049] In practice, as described above, the cloud of points NP is defined in three dimensions, represented, in this example, in an orthogonal coordinate system (U, V, W), as shown in [Fig.4].

[0050] An anonymization system 2 is described comprising a single video surveillance camera 3 and a single detection device 4; however, it is understood that the anonymization system 2 could just as well comprise a plurality of video surveillance cameras 3 and / or a plurality of detection devices 4.

[0051] In this respect, in one embodiment, the anonymization system 2 comprises a plurality of detection devices 4, each detection device 4 being of a different type, for example LIDAR, radar or infrared, which makes it possible to ensure the detection of the individual 9.

[0052] In the embodiment in which the anonymization system 2 comprises a plurality of CCTV cameras 3, 3*, as illustrated in [Fig. 8], each CCTV camera 3, 3* is configured to acquire a raw image IMb, IMb* of the same scene S. This ensures the detection of an individual 9, as will be described in more detail later. This is particularly advantageous when the CCTV cameras 3, 3* observe the same scene S from different angles.

[0053] As illustrated in [Fig. 2], the anonymization system 2 comprises a computer 5 connected to the video surveillance camera 3 and the detection device 4. In this example, the video surveillance camera 3 and the detection device 4 are connected to the computer 5 by a data transfer cable. It is understood that the video surveillance camera 3 and the detection device 4 could alternatively be connected to the computer 5 wirelessly, for example via Bluetooth, a Wi-Fi network, 5G, etc.

[0054] In this example, with reference to [Fig. 7], the computer 5 is configured to receive as input the raw IMb image from the video surveillance camera 3 and the NP point cloud from the detection device 4, and to provide as output an anonymized image IMa. The anonymization system 2 according to the invention is advantageous because it allows the anonymized image IMa to be used with few constraints, in particular, to allow its consultation by non-sworn operators, to allow its temporary or permanent storage, or to allow subsequent processing, for example, on remote and / or third-party servers to implement artificial intelligence algorithms. Thanks to the method according to the invention, a raw IMb input image becomes easily usable thereafter, as it complies with current regulations on the protection of privacy.

[0055] To this end, the computer 5 is configured to implement an anonymization process which will now be described. In this example, the computer 5 includes a computer processor which can take various forms, in particular, an FPGA or a graphics card.

[0056] Figure 7 schematically represents an implementation of an anonymization process according to the invention.

[0057] In a first step El, the computer 5 receives the raw image IMb from the video surveillance camera 2 and the point cloud NP from the detection device 4.

[0058] In a second step E2, the computer 5 identifies a sub-image IMs representative of an individual 9 in the raw image IMb, also shown in [Fig. 3]. If there are a plurality of individuals 9 in the raw image IMb, the computer 5 identifies a plurality of sub-images IMs, each sub-image IMs being representative of an individual 9. In this example, the identification is performed by the computer 5 using a learning algorithm, known as "machine learning", capable of identifying a sub-image IMs representative of an individual 9 from a database of profiles representative of an individual.

[0059] The computer 5 then determines, in the raw image IMb (and therefore in the scene S), a first position PI of the sub-image IMs. In this example, the computer 5 determines, in the (X, Y) plane, a first coordinate XI along a first axis X (corresponding in this example to a horizontal axis) and a second coordinate Y1 along a second axis Y (corresponding in this example to a vertical axis). In this example, the first position PI corresponds to the center of gravity of the sub-image IMs. It goes without saying that this could be different. It could also be the center of gravity of a box encompassing the sub-image IMs, in particular, a rectangle. Preferably, the first position PI is known in absolute or relative terms.

[0060] The first position PI of the sub-image IMs in the raw image IMb (and consequently in the scene S) determined by the computer 5 thus corresponds to a point having the coordinates, in the (X, Y) plane, of the first coordinate XI and the second coordinate Yl.

[0061] In a third step E3, the computer 5, which has received the point cloud NP from the detection device 4, determines a sub-cloud of points NPs representative of an individual 9. In this example, the identification is performed by the computer 5 using a machine learning process known as "clustering," which groups data strings by distance or similarity. Such an "unsupervised" learning process is known to those skilled in the art for data analysis.

[0062] The calculator 5 then determines a second position P2 of the subcloud of points NPs, also shown in [Fig. 5]. In this example, the calculator 5 determines, in the (X, Y, Z) coordinate system (Z being an axis orthogonal to the (X, Y) plane defined previously), a first coordinate X2 along the first X axis, a second coordinate Y2 along the second Y axis, and a third coordinate Z2. In this For example, the second position P2 corresponds to the center of gravity of the subcloud of points NPs. It goes without saying that this could be different. It could also be the center of gravity of a box encompassing the subcloud of points NPs, in particular, a rectangular prism. Preferably, the second position P2 is known in absolute or relative terms.

[0063] In a fourth step E4, the calculator 5 then determines a gap EP between the first position PI of the sub-image IMs representing an individual 9 and the second position P2 of the sub-cloud of points NPs representing an individual 9. Preferably, only the first coordinate X2 along the first X axis and the second coordinate Y2 along the second Y axis of the second position P2 are taken into account.

[0064] The computer 5 detects, in a step E5, the presence of an individual 9 in the scene S when the gap EP between the first position PI and the second position P2 is less than a predetermined threshold SI. Preferably, the predetermined threshold SI is less than 20 cm, preferably 10 cm.

[0065] Thus, for example, in the event of poor visibility of the video surveillance camera 3, due, for example, to insufficient light, the computer 5 identifies a sub-image IMs representing only a partial individual. The sub-image IMs alone does not allow the detection of the presence of the individual 9. The sub-point cloud NPs representing an individual 9, superimposed and compared to the sub-image IMs, ensures the detection of the individual 9. This is particularly advantageous when the video surveillance camera 3 and the detection device 4 are observing the same scene from different angles.

[0066] In a sixth step E6, when the deviation EP is less than the predetermined threshold SI, the computer 5 anonymizes the sub-image IMs of the raw image IMb so as to obtain an anonymized image IMa (shown in [Fig. 6]). In this example, the anonymization is performed by the computer 5 using an anonymization algorithm that allows, for example, the individual or at least their face to be blurred.

[0067] In one embodiment, the computer 5 also includes an algorithm for recognizing the colors and / or reflectivity of images and point clouds. In this example, the computer 5 determines a first colorimetry Cl of the sub-image IMs and a second colorimetry C2 of the sub-point cloud NPs. Alternatively or complementarily, the computer 5 could determine a first reflectivity RI of the sub-image IMs and a second reflectivity R2 of the sub-point cloud NPs. The computer 5 then determines a difference EC between the first colorimetry Cl of the sub-image IMs and the second colorimetry C2 of the sub-point cloud NPs and detects or confirms the presence of an individual 9 in the scene S when the difference EC between the first colorimetry Cl and The second colorimetric value C2 is below a second predetermined threshold S2. Similarly, the calculator 5 determines an ER difference between the first reflectivity RI of the sub-image IMs and the second reflectivity R2 of the sub-point cloud NPs, and detects or confirms the presence of an individual 9 in the scene S when the ER difference between the first reflectivity RI and the second reflectivity R2 is below a third predetermined threshold S3. This strengthens the comparison between the sub-image IMs and the sub-point cloud NPs and ensures the reliability of the detection of an individual 9.

[0068] In one example, as shown in [Fig. 7], the computer 5 also determines a first shape Fl of the sub-image IMs and a second shape F2 of the sub-point cloud NPs. The term "shape" refers to the detection of a contour of the sub-image IMs or the sub-point cloud MPs. The computer 5 then determines a difference between the first shape Fl and the projection of the second shape F2 in two dimensions, for example, and increases the reliability of detecting one or more individuals if the difference is less than a predetermined threshold. The difference detection can be based on a criterion of area, width, height, or another geometric parameter. The first shape Fl and / or the second shape F2 could alternatively be compared to a database of characteristic shapes of an individual, for example.

[0069] The identification of the first form Fl and / or the second form F2 could also be complemented by a comparison of criteria for example of size or build, so as to make the detection of an individual even more reliable and robust.

[0070] It is understood that the computer 5 could alternatively compare only the first shape identified in the sub-image IMs and the second shape identified in the sub-point cloud NPs, without the first position PI and the second position P2 being identified and compared.

[0071] It goes without saying that additional criteria could be added to increase the robustness of the detection.

[0072] In one embodiment, with reference to [Fig. 8], the anonymization system 2 comprises a plurality of CCTV cameras 3, 3* (in this example, two CCTV cameras 3, 3*). Each CCTV camera 3, 3* acquires a raw image IMb, IMb* of the same scene S. For example, the CCTV cameras 3, 3* acquire a raw image IMb, IMb* from different viewing angles. In this example, the computer 5: • identifies a sub-image IMs, IMs* representative of an individual 9 in each raw image IMb, IMb*, • determines in each raw image IMb, IMb* (and therefore in the scene S), respectively a first position PI, PI* of each sub-image IMs, IMs*, • determines an EP gap between the first positions PI, PI* of each sub-image IMs representing an individual 9 in each raw image IMb, IMb*, • detects the presence of an individual 9 in scene S when the EP deviation is less than a predetermined threshold SI, and • anonymizes each raw IMb, IMb* image to obtain an anonymized IMa, IMa* image.

[0073] The anonymization method according to the invention, through the combination of the video surveillance camera and the detection device, makes it possible to detect the presence of an individual in a scene without the manual intervention of a sworn officer. It is also possible to repeat the detection, and therefore the anonymization, on a large number of raw images from multiple video surveillance cameras. This is particularly advantageous in an autonomous train, which requires the use of a large number of cameras to function optimally. The anonymization method also ensures the detection of an individual present in a scene, even when visibility is reduced, for example, in darkness.

Claims

Demands

1. An anonymization system (2) for at least one raw image (IMb) of a scene (S) containing at least one individual (9), the anonymization system (2) comprising: • at least one CCTV camera (3) configured to acquire a raw image (IMb) representative of the scene (S), • at least one detection device (4) configured to determine a point cloud (NP) representative of the scene (S), • a computer (5) configured to: • identify at least one sub-image (IMs) representative of an individual (9) in the raw image (IMb), the sub-image (IMs) having a primary position (PI) in the scene (S), • identify at least one sub-point cloud (NPs) representative of an individual (9) in the point cloud (NP), the sub-point cloud (NPs) having a secondary position (P2) in the scene (S), • anonymize the sub-image (IMs) of the raw image (IMb) so as to obtain an anonymized image (IMa) if a deviation (EP),determined between the primary position (PI) and the secondary position (P2), is less than a predetermined threshold (SI).

2. Anonymization system (2) according to claim 1, wherein the detection device (3) is a remote sensing sensor of the LIDAR, radar, infrared type.

3. Rail vehicle (1) comprising at least one anonymization system (2) according to any one of claims 1 to 2.

4. A method for anonymizing at least one raw image (IMb) of a scene (S) comprising at least one individual (9), the method being implemented by an anonymization system (2) according to any one of claims 1 to 2, the method comprising the steps of: • identifying at least one sub-image (IMs) representative of an individual (9) in the raw image (IMb) acquired by a CCTV camera (3), the sub-image (IMs) having a primary position (PI) in the scene (S), • identify at least one sub-point cloud (NPs) representative of an individual (9) in a point cloud (NP) determined by a detection device (4), the sub-point cloud (NPs) having a secondary position (P2) in the scene (S), • determine a gap (EP) between the primary position (PI) and the secondary position (P2), and • anonymize the sub-image (IMs) of the raw image (IMb) so as to obtain an anonymized image (IMa) if the gap (EP) is less than a predetermined threshold (SI).

5. An anonymization method according to claim 4, the method comprising the steps of: • identifying a first colorimetry (Cl) and / or a first reflectivity (RI) of the sub-image (IMs), • identifying a second colorimetry (C2) and / or a second reflectivity (R2) of the sub-point cloud (NPs), and • determining a gap (EC, ER) between the first colorimetry (Cl) and / or the first reflectivity (RI) and the second colorimetry (C2) and / or the second reflectivity (R2).

6. An anonymization method according to any one of claims 4 and 5, the method comprising a step of: • identifying a first shape (F1) in the sub-image (IMs), • identifying a second shape (F2) in the sub-point cloud (NPs), and • determining a difference between the first shape (F1) and the second shape (F2).

7. A computer program-type product comprising at least one sequence of instructions stored and readable by a computer and which, once read by that computer (5), causes the execution of the steps of the process according to any one of claims 4 to 6.

8. Computer-readable medium containing the computer program product according to claim 7.

Images