Method for detecting flashing lights

Abstract

A method for detecting flashing lights, implemented by a device comprising at least one image acquisition device and a computer, characterized in that it comprises: an acquisition (100) of an image stream; for at least one image of said image stream, a segmentation (120) of the image comprising: a calculation (121) of a differential image resulting from the pixel-by-pixel difference in intensity between the image under consideration and a preceding image in the stream, as well as an extraction (122) of the segments of the differential image exhibiting absolute values ​​of intensity difference greater than a predefined threshold; and for each of the segments of said segmentation, an association (130) of the segment with a flashing light tracking marker in the image stream. Abstract figure: Figure 1

Description

Title of the invention: Method for detecting flashing lights Technical field

[0001] The present disclosure relates to the field of image processing. The invention relates more particularly to a method for detecting flashing lights. Prior art

[0002] When a driver takes to the road with his vehicle, he may encounter unusual situations. These situations, which are outside the driver's daily routine and habits, are often sources of dangerous and unpredictable behavior. To attract the attention of drivers, the highway code includes the use of flashing warning lights.

[0003] For example, such warning devices are found on fixed elements of the environment such as road traffic lights (when the yellow light flashes to warn of a dangerous situation, the intersection), alongside certain traffic signs (in particular for dangerous areas of encounter with pedestrians), or in work zones to draw attention to temporary signage. But they are also found on mobile elements such as on any vehicle with indicators, vehicles in emergency intervention with the rotating beacon of police cars, ambulances, or fire engines, or simply in intervention, such as the rotating beacon of agricultural machinery.

[0004] In the prior art, it is known to equip vehicles with devices capable of recognizing elements of the roadway, such as signaling elements or vehicles, in particular to provide road users with tools enabling them to improve safety, and in particular by equipping autonomous or semi-autonomous vehicles in order to improve their efficiency.

[0005] Methods for detecting potential light sources are thus known in the art, based on a segmentation of an image on the basis of a pixel intensity threshold, and depending on the applications, the intensity of certain color components of the pixels relative to other components. These methods, however, have certain limitations: they are only sufficiently effective with lights of constant intensity and / or lights of fixed position, such as red lights.

[0006] When the lights are flashing and / or follow different intensity variation patterns, these methods are inefficient and generate a large number of errors. In addition, when they are mobile, it is very difficult to maintain tracking of the flashing lights when they are in an off phase. In addition, the capture of a distant flashing light sees its saturation decrease, making the methods ineffective. methods based on the intensity of a color component of the pixels and those based on the light intensity of the pixels.

[0007] The methods of the prior art are therefore not sufficient to effectively detect flashing lights, for example of emergency vehicles, while they are generally used to prevent dangerous situations. There is a need for a solution that covers these situations. Summary

[0008] The present disclosure improves the situation.

[0009] A method for detecting flashing lights is proposed, implemented by a device comprising at least one image acquisition device and a computer, further comprising: - an acquisition of an image stream comprising a plurality of successively acquired images forming said image stream, - for at least one image of said image stream, • an image segmentation including: • a calculation of a differential image resulting from the difference in intensity, pixel by pixel, between the image considered and an image which precedes it in the flow, • extraction of segments of the differential image presenting absolute intensity difference values ​​greater than a predefined threshold, • for each of the segments of said segmentation, an association of the segment with a flashing light tracking marker in the image stream, said tracking marker being determined as follows: • if there is a tracking marker at a distance from the segment considered less than a predefined threshold, said segment is associated with said tracking marker, and the position of said marker is updated to that of the segment, and • otherwise, a tracking marker is defined in the image stream at the position of the segment in question and is associated with it.

[0010] In a variant of the method for detecting flashing lights, the association of a segment with a tracking marker further comprises a verification that the sequence of segments associated with the tracking marker, including the segment in question, respects a predefined pattern, for example an alternation of rising and falling edges of intensity of the segments.

[0011] In another variant, the method for detecting flashing lights further comprises a classification of the tracking markers.

[0012] Alternatively, tracking markers are classified into different types of lights.

[0013] Optionally, the different types of lights include lights flashing in different patterns, different frequencies, and different intensities, as well as steady lights.

[0014] In a variant of the method for detecting flashing lights, the classification of each tracking marker is carried out by analyzing a pattern formed by the sequences of segments associated with the marker

[0015] Alternatively, the classification of each tracking marker is performed by a machine learning classification algorithm, pre-trained on a training database comprising a plurality of image streams of different types of lights such as lights flashing in different patterns, different frequencies, and different intensities.

[0016] In one embodiment, the method for detecting flashing lights further comprises, for each acquisition of an image from said image stream: - a calculation of the displacement, of an image considered relative to the image which precedes it in the stream, of elements present in said images, - updating the position of at least one tracking marker, from the calculated displacement.

[0017] In a variant, the method for detecting flashing lights further comprises removing tracking markers that have not been associated with any segments extracted from a number of images greater than a predefined limit.

[0018] In one embodiment, the method for detecting flashing lights further comprises, for each acquisition of an image of the flow: - a definition, prior to said segmentation, of an area of ​​interest on the image considered, and - implementation of image segmentation only on the area of ​​interest,

[0019] In one embodiment, the image acquisition device is on board a vehicle, and the area of ​​interest is determined from the points at infinity and vanishing lines of the image, such as the area comprising the elements substantially fixed from one image to another of the image stream.

[0020] In a variant of the method for detecting flashing lights, the definition of the area of ​​interest further comprises: - the evaluation of the distance of certain elements present in the image, to the device acquiring said image, and - the exclusion of areas of the image containing elements at a distance thus assessed as less than a predefined limit distance.

[0021] Alternatively, the method of detecting flashing lights further comprises dynamic object detection and tracking and the area of ​​interest includes pixels associated with detected and tracked dynamic objects.

[0022] According to another aspect, there is provided a computer program product, comprising code instructions for implementing the method of detecting flashing lights, when executed by a computer.

[0023] According to another aspect, there is provided a device for detecting flashing lights comprising an image acquisition device and an on-board or remote computer configured to implement the method for detecting flashing lights.

[0024] According to a final aspect, a motor vehicle is proposed comprising a device for detecting flashing lights. Brief description of the drawings

[0025] Other characteristics, details and advantages will appear on reading the detailed description below, and on analyzing the attached drawings, in which: Fig.l

[0026] [Fig.l] [Fig.l] shows a schematic representation of the progress of the method according to one embodiment. Fig. 2

[0027] [Fig.2] [Fig.2] shows a schematic representation of a vehicle implementing the method according to one embodiment. Fig. 3

[0028] [Fig.3] [Fig.3] schematically shows successive illustrations of the different stages of the process, according to one embodiment of the latter. Description of the embodiments

[0029] For example, in order to provide a safety tool or to make it autonomous for detecting dangerous situations, a vehicle must be able to recognize the flashing lights that it encounters on the road. Reference is now made to [Fig. 1], which schematically represents a flowchart of the progress of the method for detecting flashing lights according to one embodiment, and to [Fig. 2] which represents a vehicle according to one embodiment of the method.

[0030] This method is implemented by a device 200 for detecting flashing lights, comprising at least one computer 201, for example of the processor, microprocessor, microcontroller, programmable logic circuit (FPGA, etc.) type, and / or a specialized integrated circuit (ASIC, etc.). The computer 201 is advantageously embedded in the vehicle 210. Alternatively, the computer 201 may be located remotely from the vehicle and be in remote communication with an on-board communication device 203 and / or a component of the vehicle 204.

[0031] For example, the communication device 203 is a device following the V2X (Vehicle-to-everything) vehicular communication technology and particularly the V2N (Vehicle-to-Network) technology.

[0032] The device may further comprise a memory 202, storing code instructions executed by the computer 201 to be able to implement the method. The memory 202 may also store one or more pre-trained models described in more detail below to be able to determine in more detail the types of lights detected. The memory 202 may be a non-volatile electronic memory, an optical hard disk, etc.

[0033] The vehicle 210 further comprises an image acquisition device 204, making it possible to provide the computer 201 with images for implementing the method. This device 204 may be an on-board camera or an on-board camera.

[0034] In order to study the flashing lights of the vehicle's environment, the method firstly comprises, in one embodiment, an acquisition 100 of an image stream. This acquisition 100 can be carried out by means of the image acquisition device 204 on board the vehicle, for example from a video stream or a series of successive photos.

[0035] In a variant, this acquisition 100 comprises a pre-processing of the images, depending on the application case. Certain components of the pixels can be deleted or undergo a modification, so that the images are advantageously adapted to the application case. For example, to detect only red flashing lights, the green and blue components of the pixels of the images are deleted. In another example, the images are advantageously modified into shades of gray to reduce memory and calculations.

[0036] Acquiring such a stream of images advantageously makes it possible to observe an evolution over time of certain elements present in the image. In particular, the flashing of a light is a phenomenon which spreads over time.

[0037] Thus, the method comprises comparing an image of the stream with an image that precedes it in the stream. The image that precedes it is not necessarily the one that immediately precedes it in the stream; several other images may separate them. Furthermore, the stream of images may be previously sub-sampled, so as to, advantageously, not contain images that separate the images to be compared.

[0038] This comparison is carried out by the calculation 121 of a so-called differential image, because it results from the difference, pixel by pixel, between the two images. Such a differential image expresses the variations that each pixel has undergone during the time which separates the taking of each of the images. Consequently, it advantageously contains the information of rising and falling edges of flashing lights.

[0039] The differential image Imdiff can thus be constructed in the following manner:

[0040] Imdiff ~ with 'a nth image considered in the stream and k the number of images which separates the image considered and the one with which it is compared and which precedes it in the flow.

[0041] Generally, the acquisition 100 of the images of the flow by such a device 204 follows a constant acquisition period, and we can then also note:

[0042] Imdiff = Imt - with t the time at which the image considered is acquired and T the acquisition period.

[0043] The difference between two images is understood to mean the difference, for each of the pixels of the first image, between the components of the pixel in question and the respective components of the corresponding pixel (at the same position) in the second image. The result of such a difference can advantageously result in positive as well as negative component values ​​for the pixels of the differential image. For example, if the red component of a pixel is negative, this indicates a decrease in the red level for this pixel during the time separating the acquisition of the two images.

[0044] The components or color components are the values ​​that allow the color of a pixel to be expressed, for example the quantity of red, green and blue, or the hue, saturation and vibrance. In variants, this can also correspond to the gray level of the pixel, its luminance, etc.

[0045] The differential image can thus express the variations in intensity of the pixels. The variations in intensity considered can depend on the application case and can include: variations in the overall intensity of the pixel, that of one or more of its color components such as a color variation from red to blue, or a combination of variations in the overall intensity and color of the pixel.

[0046] The number of images k which separates the image Imn from the image Imnk can be constant, for example equal to 1, so as to compare two images acquired consecutively.

[0047] In a variant, the number k can also follow a variable function, for example a triangular or sawtooth function. Indeed, for a constant k, a flashing light with a period multiple of kT is in phase with the calculation 121 of the differential image: the difference is then always made between two images on which the flashing light is in the same state (either always on or always off) and substantially no variation is then measured on the differential image. A number of images of variable separation k advantageously makes it possible to eliminate this risk.

[0048] In one embodiment, the calculation of the differential image comprises a realignment of the image Imn with respect to the image Imnk taking into account the displacement from one image to the other of elements present in both images.

[0049] The method then comprises the extraction 122 of the areas of the differential image, called segments, when the absolute intensity values ​​of the pixels of the area are greater than a predefined threshold. For example, the threshold is a variation of + / - 20%. The predefined value for the threshold can advantageously be adapted to each application case.

[0050] Depending on the application, the intensity of a segment is understood to mean an average (or a median, etc.) of the intensity of the pixels that make up this segment. This intensity may relate to a particular component of the pixels of the segment or a combination of components, or the result of a calculation on a component or a combination of components, such as for example the luminance, saturation, or gray level of the pixels of the segment.

[0051] In a variant, the extracted area is approximated by a geometric shape, for example elliptical or preferably rectangular, so as to be able to advantageously easily define the position of a segment by a remarkable point (center or corner) rather than by a barycenter of the pixels of any shape.

[0052] The segmentation 120 of the method, which comprises the calculation 121 of the differential image and the extraction 122 of the segments, makes it possible to establish a set of segments which each correspond to a strong variation (defined by the threshold), and which are thus potential candidates for being the witness of the presence of one or more flashing lights in the image.

[0053] In one embodiment, the method advantageously comprises, prior to the segmentation 120, the definition 110 of an area of ​​interest on the image, to which the segmentation 120 is limited.

[0054] Indeed, the extracted segments correspond to areas whose pixels have undergone significant variations. However, when the vehicle is moving with the on-board image acquisition device 204, the closer an element of the Imnk image was physically to the vehicle, the more likely it is to have a different position on the Imn image. This difference in position can lead to major variations in the pixels and therefore an extraction of false-positive segments associated with these variations.

[0055] Defining an area of ​​interest in the image, not including elements too close to the vehicle, to limit the segmentation advantageously makes it possible to avoid extracting segments corresponding to the movement of nearby elements. The segmentation thus advantageously has better recall (or better sensitivity).

[0056] The definition of the area of ​​interest is generally carried out by applying a filter to the images, so as to apply a multiplicative weight (or coefficient) of 0 to the (components of) pixels outside the area of ​​interest and of 1 to the pixels inside. of the area of ​​interest.

[0057] In a variant not illustrated, said filter comprises weights between 0 and 1, so the delimitation of the area of ​​interest is not clear. For example, the area of ​​interest can extend around a point of interest, so that the further a pixel is from the point of interest, the lower its weight, following a linear, polynomial, exponential or logarithmic decrease, etc.

[0058] In one embodiment, the area of ​​interest may be defined as including a set of points considered at infinity relative to the image acquisition device.

[0059] The delimitation of the area of ​​interest on the image can be done from the vanishing lines (and the vanishing point or point at infinity which can be deduced therefrom) which advantageously make it possible to determine the relative distance between elements of the image with respect to the image acquisition device at the origin of the image. In particular, a set of points located at a distance less than a predefined distance threshold from the vanishing point of the image can be approximated as points located at infinity.

[0060] Generally, the vanishing point is located in the central part of the image, towards which the vanishing lines are oriented. The delimitation of the area of ​​interest can thus be fixed (identical, whatever the image) beforehand, like the set of pixels in the central part. For example, it can be in the form of an oval encompassing all the pixels in the central part. Such a delimitation advantageously makes it possible to reduce the calculation time necessary for defining the area of ​​interest.

[0061] However, the area of ​​interest can also be more or less variable depending on the image to which it is applied. In a variant, the area of ​​interest is delimited from a predefined threshold of distance of the elements of the image relative to the image acquisition device: the elements located at a distance less than the predefined threshold delimit the area of ​​interest. Such a delimitation makes it possible to have an area of ​​interest that is very faithful to reality. In addition, the computer 201 can advantageously perform recognition of the 3D geometry of the object in order to specify the evaluation of its distance from the image acquisition device 204.

[0062] In one variant, the computer 201 detects and tracks dynamic objects in the image stream. For example, the computer implements a Kalman filter on the images. The area of ​​interest may advantageously include the pixels of the dynamic objects, so that, for example, vehicles likely to emit flashing lights such as emergency vehicles are part of the area of ​​interest.

[0063] In another variant, the area of ​​interest is of a predefined shape, for example, preferably an ellipse, the geometric characteristics of which (in the example, the semi-major axis, the semi-minor axis and the center) can vary according to the image or the pair of images. The center of the ellipse can advantageously be the vanishing point of the image.

[0064] The dimensions of the area of ​​interest can also take into account the movement made by the vehicle which equips the image acquisition device, between the images or the speed of the vehicle. For example, the slower the vehicle 210 moves between the two images, the wider the area of ​​interest can be.

[0065] Thus, this latter variant advantageously makes it possible to obtain an effective area of ​​interest, without requiring too many computing resources and time.

[0066] Alternatively, in this latter variant or in a variant where the area of ​​interest is previously fixed, the delimitation comprises a subsequent verification. This verification excludes elements from the foreground of the image, both close to the image acquisition device 210 and close to the vanishing point on the image. For example, this element is a vehicle which follows the vehicle 210 equipped with the image acquisition device 200. This verification advantageously makes it possible to exclude elements which could initially be included in the area of ​​interest by their proximity on the image to the vanishing point, despite their proximity to the device 210. These elements can be detected by means of known methods for detecting road elements and in particular for detecting vehicles.

[0067] The method then comprises the association 130 of the segments with markers for tracking a flashing light. A tracking marker makes it possible to maintain the tracking of a flashing light from one image to another of the flow. Two cases are possible for a segment thus extracted during the segmentation: either a pre-existing tracking marker is close (i.e. less than a predefined distance threshold depending on the application case) to this segment, or the segment is isolated (no tracking marker is close to it).

[0068] An isolated segment corresponds to a sudden variation (i.e. a rising or falling edge) of the intensity of pixels in a part of the image which had not previously varied: it is a candidate for the appearance of a new flashing light in the image. Consequently, the association 130 comprises the creation of a new tracking marker, at the position of the isolated segment and associates itself with it.

[0069] A segment close to a marker is a candidate for the continuation of the flashing of a light already segmented in previous images. The method therefore advantageously comprises the association of this segment with this marker. If a segment is close to several markers, it is associated with the closest marker. When a segment is associated with a pre-existing marker, the association comprises the updating of the position of the tracking marker on the position of the segment which is associated with it, so that a marker successively associated with a sequence of segments coming from a moving light can synchronize its position with that of the moving light.

[0070] In one variant, the association includes a verification that the associated segment fits into a predefined pattern corresponding to a flashing. For example, we check that a marker previously associated with a segment linked to a light that turns on (positive sign of intensity variation in the components of the pixels of the segment on the differential image) is associated with a segment linked to a light that turns off (negative sign of intensity variation). The pattern is then an alternation of rising edges and falling edges of intensity of the segments of the sequence.

[0071] Alternatively, the predefined pattern may also relate to the intensity of the associated segments, so that the evolution of the intensity follows a flashing pattern.

[0072] In one embodiment of the method, the position of the markers associated with no segment is updated 140. A moving flashing light in the image stream continues to move when it does not generate variations / segments (stable intensity). Thus, the update 140 comprises the calculation of the displacement of the elements of the image Imn relative to the image Imnk, and from this calculation, an estimation of the displacement of the flashing light linked to the tracking marker.

[0073] These elements may be characteristic elements of the roadway present in both images, such as a lamppost that has moved from one image to the other. Alternatively, the elements may be a simple characteristic arrangement of pixels of particular intensities, when this arrangement is present in both images at different positions.

[0074] Taking into account a plurality of elements advantageously makes it possible to estimate a precise displacement.

[0075] The position of the tracking marker is thus updated by following this movement.

[0076] In a variant, the estimated displacement is zero because the images Imn and Imnk are identical and the vehicle is stationary. Advantageously, updating the tracking markers is avoided in this particular case.

[0077] Alternatively, this adjustment also takes into account a plurality of previous images, making it possible to estimate, advantageously more precisely, the displacement by correlating it with the displacements of the previous images.

[0078] Alternatively, the method may comprise detecting image elements using known methods, and advantageously updating 140 tracking markers relative to the detected elements. For example, a tracking marker corresponding to a flashing light on a detected vehicle may advantageously be updated 140 according to the movement made by the detected vehicle.

[0079] Once the marker position is updated by following the estimated displacement, the marker is theoretically at the position of the light of stable intensity. Thus, if the light generates a new segment on a following image, the correct tracking marker (corresponding to the same flashing light) will advantageously already be po located close to the segment.

[0080] Advantageously, this update allows a marker to follow a flashing light, including when it is in a stable off phase.

[0081] Since a marker associated with no segment for a long period means the disappearance of a flashing light, the updating 140 of a marker advantageously comprises its deletion. Thus, according to one embodiment of the method, a marker associated with no segment for a predefined number of images or time is deleted.

[0082] Advantageously, in a variant, this predefined number of images (or this time) corresponds to a time greater than the flashing period of the predefined pattern of the verification of the method, so that a marker is not deleted by mistake for a flashing light whose flashing period is too high.

[0083] In another variant, the method comprises deleting a marker which, once its position is updated, is located outside the area of ​​interest for a predefined number of images or time. Since no segment is created outside the area of ​​interest, the marker corresponds to a flashing light which leaves the field of the image. The predefined time advantageously corresponds to a waiting period, during which the flashing light which leaves the image can change direction and re-enter the area of ​​interest.

[0084] In one embodiment, the method comprises a classification 150 of the tracking markers which determines the type of light to which the tracking marker is linked. This classification can advantageously differentiate the lights by their physical characteristics such as their flashing period, their flashing pattern, the light frequency(ies) (colors) emitted, etc.

[0085] Such an analysis for the classification 150 also advantageously makes it possible to detect a false positive, when a flashing light tracking marker has been associated with a non-flashing light.

[0086] In a variant, such a classification 150 is carried out by studying the sequence of associations of segments with the marker studied. The method thus deduces from the sequence of associations of segments with the tracking marker, the flashing pattern of the light associated with the marker and the rest of the physical characteristics of the associated light.

[0087] In a variant of the embodiment, this classification 150 differentiates the lights by their application. This classification 150 is advantageously more utilitarian and because it makes it possible to determine which event the light corresponds to. For example, the classification determines whether the marker follows a traffic light, an ambulance flashing light, a police flashing light, etc.

[0088] Advantageously, in a variant, such a classification 150 is carried out by a Machine learning algorithm pre-trained on a set of labeled tracking markers (i.e. the type of light related to the marker is known). For example, the training database includes tracking markers of flashing lights from traffic lights, rotating beacons from different emergency vehicles, etc.

[0089] Alternatively, the training database for classification 150 also includes flashing lights from different countries so that the algorithm can be advantageously effective regardless of the country of origin of the lights. Such an algorithm is for example capable of differentiating the light coming from a French ambulance flashing light from an American ambulance flashing light.

[0090] In one embodiment, the classification 150 is carried out by the on-board computer 201 on the vehicle. Alternatively, it is carried out by a remote computer; in this case, the vehicle comprises a communication device 203 which transmits to a remote server 213 the information necessary for the classification (the set of tracking markers) and the remote server 213, after having carried out the classification calculation, transmits the results to the vehicle.

[0091] Reference is now made to [Fig.3] which illustrates the implementation of the different stages of the method according to one embodiment.

[0092] In this embodiment, the computer 201 receives the image 300 from the image acquisition device 204. This image comprises an ambulance whose first rotating light 301 has just turned on, and the second rotating light 302 remains off (i.e. it was already off during the previous segmentation and still is).

[0093] Image 310 illustrates the definition 110 of an area of ​​interest 311 from the vanishing lines present in image 300. In this application case, a fixed elliptical shape and size are predefined, and the ellipse is centered on the vanishing point (at infinity) of the image.

[0094] Furthermore, this definition includes the exclusion of the area 312 comprising an element of the foreground of the image (a car) located both close to the vanishing point of the image and physically close to the image acquisition device 204.

[0095] Image 320 is the differential image resulting from the difference 121 of image 300 and that which precedes it (not illustrated) in the image stream of the image acquisition device 204. The black zones represent small variations while the white zone 321 (located at the level of the first rotating beacon 301) represents a zone of large variations in intensity (in absolute value).

[0096] Image 330 thus illustrates the result of segmentation 120 of image 300. A new segment 333 is extracted 122 from area 321. This segment 333 is associated 130 with the first marker 332a, which is the marker 332 closest to segment 333.

[0097] Since the segment corresponds to a light turning on, the method also includes verifying that the last segment associated with marker 332a was a segment corresponding to a light turning off, in order to follow a flashing pattern with the new segment 333.

[0098] The second marker 332b, resulting from a previous segmentation (of the second rotating beacon 302 of the ambulance), is not associated with any segment. Its position is nevertheless updated 140 in order to be located at the position where the light of the second rotating beacon of the ambulance is estimated to reappear.

[0099] The method then comprises a classification 150 of the marker 332a, which, by its pattern of associations with the segments resulting from previous segmentations, is recognized as linked to an ambulance flashing light.

Claims

Claims

1. Method for detecting flashing lights, implemented by a device (200) comprising at least one image acquisition device (204) and a computer (201), characterized in that it comprises: an acquisition (100) of an image stream comprising a plurality of successively acquired images forming said image stream, for at least one image of said image stream, • a segmentation (120) of the image comprising: • a calculation (121) of a differential image resulting from the difference in intensity, pixel by pixel, between the image considered and an image which precedes it in the flow, • an extraction (122) of the segments of the differential image presenting absolute values ​​of intensity difference su above a predefined threshold, for each of the segments of said segmentation, an association (130) of the segment with a flashing light tracking marker in the image stream, said tracking marker being determined as follows: • if there is a tracking marker at a distance from the segment considered less than a predefined threshold, said segment is associated with said tracking marker, and the position of said marker is updated (140) to that of the segment, and • otherwise, a tracking marker is defined in the image stream at the position of the segment in question and is associated with it.

2. A method of detecting flashing lights according to claim 1 wherein the association (130) of a segment with a tracking marker further comprises a verification that the sequence of segments associated with the tracking marker, including the segment in question, respects a predefined pattern.

3. A method of detecting flashing lights according to any preceding claim further comprising classifying (150) the tracking markers.

4. A method of detecting flashing lights according to the preceding claim, wherein the tracking markers are classified among different types of lights.

5. A method of detecting flashing lights according to the preceding claim, wherein the different types of lights include lights flashing in different patterns, different frequencies, and different intensities, as well as steady lights.

6. A method of detecting flashing lights according to any one of claims 3 to 5, the classification (150) of each tracking marker is carried out by analyzing a pattern formed by the sequences of segments associated with the marker.

7. A method for detecting flashing lights according to any one of claims 3 to 6, wherein the classification (150) of each tracking marker is performed by a machine learning classification algorithm, pre-trained on a training database comprising a plurality of image streams of different types of lights such as flashing lights in different patterns, different frequencies, and different intensities.

8. Method for detecting flashing lights according to any one of the preceding claims further comprising, for each acquisition of an image from said image stream: - a calculation of the displacement, of an image considered relative to the image which precedes it in the stream, of elements present in said images, - updating the position of at least one tracking marker, from the calculated displacement.

9. A method of detecting flashing lights according to any preceding claim further comprising removing tracking markers that have not been associated with any segments extracted from a number of images greater than a predefined limit.

10. A method of detecting flashing lights according to any preceding claim further comprising, for each ac- acquisition of an image of the flow: - a definition (110), prior to said segmentation, of an area of ​​interest on the image considered, and - the implementation of the segmentation (120) of the image only on the area of ​​interest.

11. Method for detecting flashing lights according to the preceding claim in which the image acquisition device (204) is on board a vehicle (210), and the area of ​​interest is determined from the points at infinity and vanishing lines of the image, such as the area comprising the elements substantially fixed from one image to another of the image stream.

12. Method for detecting flashing lights according to any one of the preceding claims in which the definition of the area of ​​interest further comprises: - the evaluation of the distance of certain elements present in the image, to the device for acquiring said image, and - the exclusion of areas of the image comprising elements at a distance thus evaluated to be less than a predefined limit distance.

13. A method of detecting flashing lights according to claim 11 or 12 further comprising detecting and tracking dynamic objects and the area of ​​interest comprises the pixels associated with detected and tracked dynamic objects.

14. Computer program product, comprising code instructions for implementing the method according to any one of the preceding claims, when executed by a computer (201).

15. Device (200) for detecting flashing lights comprising an image acquisition device (204) and an on-board or remote computer (201) configured to implement the method for detecting flashing lights according to any one of claims 1 to 13.

16. Motor vehicle (210) comprising a device (200) for detecting flashing lights according to the preceding claim.

Images