Improvement in the detection of an attribute in an image of a camera of a vehicle

Abstract

The invention relates to a method for detecting an attribute in an image acquired by an image-acquiring device of a vehicle arranged to acquire images representative of a scene facing the vehicle, the method comprising the following steps of a current phase: - obtaining (201) at least one feature of at least one attribute detected in the image and contained in a segment of the image; - determining (202) at least one state of the image depending on the at least one obtained feature; - determining (203), for each determined state, an action depending on the determined state, by applying a model designed and parameterized to receive as input an image state and to determine as output an action dependent on the state received as input; - implementing (204) said action, the action being a modification of the image to obtain a modified image, or an adaptation of the lighting of the vehicle.

Description

Improved detection of an attribute in an image from a vehicle camera

[0001] The present invention relates to the field of detecting attributes in an image acquired by a camera of a motor vehicle.

[0002] It is particularly advantageous for improving a detection and / or segmentation function implemented from images from a camera facing a scene located in front of the vehicle, in night situations.

[0003] Most motor vehicles now incorporate driver assistance functions to assist the driver in steering the vehicle, thus improving safety and driving comfort.

[0004] To this end, certain functions are based on the detection of attributes in a scene facing the vehicle, in particular from images captured by a vehicle camera facing the front of the vehicle, and in addition from data from other sensors, such as radar or lidar for example.

[0005] The images captured by the camera can be segmented in order to detect an attribute in each segment, a driving assistance function being able to exploit the attribute detected in a segment to implement an action improving the safety or driving comfort of the driver. Other functions can be implemented from the attributes detected in the scene, such as adaptive lighting functions for example.

[0006] In order to segment an image and detect an attribute in a segment, machine learning models are developed and implemented in software form in motor vehicles to receive images from a camera as input and to transmit detected attributes as output to at least one adaptive or driver assistance function of the vehicle.

[0007] Such models offer high performance when the external brightness, and therefore the brightness of the scene captured by the camera, is above a given threshold (in daytime driving situations, or on a well-lit road at night).

[0008] However, in some night driving situations, the performance of segmentation and attribute identification models is much lower, despite the vehicle's lighting functions. Attribute detection errors can therefore occur, posing safety problems or reducing driving comfort.

[0009] It is known to adapt the lighting in certain areas of the scene facing the vehicle. For this purpose, the vehicle may comprise lighting modules comprising a matrix light source capable of projecting a pixelated beam, the pixels of which are individually controllable.

[0010] This makes it possible to illuminate differently the areas of the environment facing the vehicle corresponding to the different segments, in order to try to improve the detection of attributes in each segment.

[0011] However, it is complex to determine, in anticipation, how to improve the lighting in a segment of the given scene to improve the contrast and / or improve the detection of attributes.

[0012] There is thus a need to improve the detection of attributes in an image representative of the scene facing the vehicle, in night driving situations.

[0013] To this end, a first aspect of the invention relates to a method for detecting an attribute in an image acquired by an image acquisition device of a vehicle, the image acquisition device being arranged to acquire images representative of a scene facing the vehicle, the method comprising the following steps of a current phase: - obtaining at least one characteristic of at least one attribute detected in the image and included in a segment of the image; - determining at least one state of the image as a function of the at least one characteristic obtained; - determining, for each determined state, an action as a function of the determined state, by applying a model designed and parameterized to receive as input an image state and to determine as output an action dependent on the state received as input; - implementing said action, the action being a modification of the image to obtain a modified image, or an adaptation of the lighting of the vehicle.

[0014] Thus, improved attribute detection can be implemented in the modified image or in a new image acquired by the vehicle camera after adaptation of the lighting, which is particularly advantageous in night driving situations in which the image initially acquired by the camera makes it difficult to identify with certainty an attribute in the scene facing the vehicle. Attribute detection is useful both for driving assistance functions of the vehicle, but also for the driver directly, who can adapt his driving to the detected attribute.

[0015] According to embodiments, the at least one characteristic obtained may comprise at least one confidence score associated with the detection of the at least one attribute and / or a local contrast value in the segment comprising the at least one attribute.

[0016] Thus, the state determined, and provided as input to the model, depends on a characteristic directly linked to an ability to identify with certainty an attribute in the object. The model can thus target as a priority, by its action or actions, one or more attributes having a confidence score or a low local contrast value.

[0017] Additionally, the model can be parameterized so that, for at least one given attribute detected in the image, a confidence score associated with the detection of the given attribute and / or a local contrast value of the at least one given attribute, is higher in the modified image or in a new image acquired by the camera after adaptation of the lighting of the vehicle, than in the image.

[0018] Thus, the model is trained / parameterized in advance to determine an action that improves the detection of at least one attribute of the image, which increases driving safety and comfort, particularly in night driving situations.

[0019] Additionally or alternatively, the at least one characteristic obtained further comprises one or more of the following characteristics:- one or more attribute characteristics from among an attribute type, an attribute position, an attribute shape and an attribute size;- a local brightness value in the segment of each detected attribute;- an overall contrast value in the image;- an average brightness value in the image.

[0020] Thus, the accuracy of the model is improved because it can take as input a large number of different states, defined by a rich set of descriptive characteristics of the image and of the at least one attribute detected there. The determined actions can thus be richer.

[0021] According to embodiments, the method may further comprise a training phase, the training phase comprising reinforcement learning of the model, from training images.

[0022] Reinforcement learning allows the training of an agent on training data, with the agent's actions being evaluated to determine a reward, positive or negative, based on which the model parameters are updated by the agent. The model thus allows the mapping between input states and respective output actions. Moreover, reinforcement learning allows for long-term strategies in determining the best action, with actions being evaluated based on the cumulative rewards they allow, and not only on the reward obtained immediately following their application.

[0023] In addition, reinforcement learning may comprise an evaluation of each action from the model from a state determined for a training image, the evaluation being based, for a given action, on:- a difference between a first initial confidence score associated with an attribute detected in the training image, and a second confidence score determined for the attribute after implementation of the action, in the modified training image or in an evaluation image received following the lighting adaptation;- a difference between a local contrast value determined for a segment comprising the attribute detected in the training image and a local contrast value determined for the segment comprising the attribute after application of the action.For each evaluation, a reward is determined, said model being updated according to said reward.

[0024] Thus, the model is specifically trained to enhance an image feature or attribute that directly impacts the ability to detect the attribute in the scene, with high certainty.

[0025] Additionally, reinforcement learning can be based on a Q-learning type algorithm.

[0026] Such an algorithm makes it possible to avoid defining initial rules, or "policies" in English, when learning the model.

[0027] According to embodiments, determining the at least one state of the image may comprise determining a local state for each attribute detected in the image, based on at least one characteristic determined for said attribute, and the model may be applied to each local state to obtain a local action, the implementation of each local action determined for an attribute comprising a modification of the segment of the image comprising the attribute to obtain the modified image or an adaptation of the lighting in an illuminated area corresponding to the attribute.

[0028] This makes it possible to improve the detection of multiple attributes in a scene facing the vehicle, which improves driving comfort and safety.

[0029] According to embodiments, each attribute detected in the image may be one of the types including a pedestrian, a road infrastructure element, an object, or other vehicle.

[0030] Thus, it is made possible to improve attribute detection for a wide range of attribute types.

[0031] According to a first embodiment, the action may be a modification of the image to obtain a modified image, the method further comprising the detection of said at least one attribute in the modified image and the transmission of the modified image to a driving assistance module of the vehicle capable of implementing at least one driving assistance function as a function of the at least one attribute detected in the modified image, and / or the transmission of the modified image and descriptive information of the at least one attribute detected in the modified image to a screen of the vehicle.

[0032] Thus, improving attribute detection allows for improved accuracy in implementing the driving assistance function, or for providing precise information to the driver. This improves driving safety and comfort.

[0033] Additionally in the first embodiment, the image modification may be a gamma transformation function defined by a gamma parameter value selected from a predetermined set of gamma parameter values, each gamma transformation function mapping an input gray level to an output gray level.

[0034] Thus, it is possible to predetermine a large number of different actions by varying a single parameter.

[0035] According to a second embodiment, the action may be an adaptation of the lighting of the vehicle, the vehicle comprising a lighting module comprising a matrix light source, the matrix light source comprising a set of individually controllable light elements, and the adaptation of the lighting may comprise a modification of a light intensity value of at least one light element capable of emitting an elementary light beam towards an illuminated zone corresponding to at least one segment comprising an attribute detected in the image.

[0036] Improving lighting to facilitate the identification of the attribute in a soon-to-be-acquired image also improves the readability of the scene facing the driver, thereby directly improving safety and driving comfort.

[0037] A second aspect of the invention relates to a computer program comprising instructions for implementing the method according to the first aspect of the invention, when these instructions are executed by a processor.

[0038] A third aspect of the invention relates to a control module for a motor vehicle, the control module being capable of obtaining at least one characteristic of at least one attribute detected in an image and included in a segment of the image, the image coming from an image acquisition device of the vehicle, the image acquisition device being arranged to acquire images representative of a scene facing the vehicle, the control module comprising a processor configured to:- determine at least one state of the image as a function of the at least one characteristic obtained;- determine, for each determined state, an action as a function of the determined state, by applying a model designed and parameterized to receive as input an image state and to determine as output an action dependent on the state received as input;- implementing said action, wherein the action is a modification of the image to obtain a modified image, or an adaptation of the lighting of the vehicle.;

[0039] A fourth aspect of the invention relates to a motor vehicle comprising an image acquisition device arranged to acquire images representative of a scene facing the vehicle, and a control module according to the third aspect of the invention.

[0040] According to the first embodiment, the action may be a modification of the image to obtain a modified image, the vehicle may comprise an image processing module capable of detecting said at least one attribute in the image and in the modified image, the vehicle further comprising a driving assistance module capable of implementing at least one driving assistance function as a function of the at least one attribute detected in the modified image, and / or a screen of the vehicle capable of receiving the modified image and descriptive information of the at least one attribute detected in the modified image, for display on said screen.

[0041] According to the second embodiment, the action may be an adaptation of the lighting of the vehicle, the vehicle comprising a lighting module comprising a matrix light source, the matrix light source comprising a set of individually controllable light elements, and the adaptation of the lighting may comprise a modification of a light intensity value of at least one light element capable of emitting an elementary light beam towards an illuminated area corresponding to at least one segment comprising an attribute detected in the image.

[0042] Other characteristics and advantages of the invention will appear on examining the detailed description below, and the appended drawings in which:

[0043] illustrates a motor vehicle according to embodiments of the invention;

[0044] illustrates the steps of a common phase of a method according to embodiments of the invention;

[0045] illustrates an example of segmentation of an image acquired by a camera of a vehicle according to embodiments of the invention;

[0046] illustrates a system for learning a model according to embodiments of the invention;

[0047] illustrates the steps of a learning phase of a method according to embodiments of the invention;

[0048] illustrates examples of actions for modifying an image, according to the first embodiment of the invention;

[0049] illustrates the structure of a control module according to embodiments of the invention.

[0050] The description focuses on the characteristics that distinguish the method and device from those known in the state of the art.

[0051] Illustrates a vehicle 100 according to embodiments of the invention.

[0052] The vehicle 100 comprises a control module 101 according to embodiments of the invention, capable of determining an action and implementing the action as described below.

[0053] The vehicle 100 further comprises an image acquisition device 102 capable of acquiring images representative of a scene facing the vehicle, i.e. in front of the vehicle in a direction of travel of the vehicle. In particular, the image acquisition device 102 may be located at the front of the vehicle, for example at the top of the windshield of the vehicle 100 or at the bumper of the vehicle 100. The image acquisition device 102 is called camera 102 in the following, for the sake of simplification. It will be understood that the camera 102 covers any device for acquiring images one by one, or in the form of frames of a video stream.

[0054] The acquired images can be color images, coded for example in red-green-blue, or RGB in English for “Red-Green-Blue”, that is to say that the first images include colored pixels arranged in a matrix, and that each pixel is coded so as to correspond to one color among strictly more than two colors. For example, each pixel can be coded in red-green-blue, or RGB for “Red Green Blue”, that is to say that the set of colors is represented by three coordinates respectively red, green and blue. Each pixel can thus be coded on several bits, for example on a byte coding 256 colors.

[0055] Alternatively, the images acquired by the camera 102 may be grayscale images, i.e. each first image comprises pixels, arranged in a matrix, each pixel being coded on strictly more than two levels, with at least one intermediate grayscale between white and black. Each pixel may thus be coded on several bits, for example on a byte coding 256 grayscale levels.

[0056] The invention applies in particular in so-called “night” driving situations, or generally when the images acquired by the camera have an average light intensity lower than a given threshold. Such a driving situation can thus be detected by a vehicle sensor (the camera 102 or another sensor) when the light intensity outside the vehicle is lower than a given, predetermined threshold, the given threshold being for example between 800 and 1200 Candelas, in particular equal to 1000 Candelas.

[0057] The vehicle 100 may further comprise an image processing module 103 capable of receiving an image, or a series of images, acquired by the camera 102. The image processing module 103 is capable of applying image processing to the received image in order in particular to segment the image into at least one segment, and to detect at least one attribute in at least one of the segments. No restriction is attached to the processing applied by the image processing module 103 to segment the image and identify attributes, which may be based on a segmentation model resulting from machine learning, which is not the subject of the present invention.

[0058] The attribute may be, for example:- a pedestrian;- another vehicle traveling in the same direction as vehicle 100;- another vehicle traveling in the opposite direction to that of vehicle 100;- a traffic sign;- the marking of the road on which vehicle 100 is traveling;- any other object or element of the road infrastructure.

[0059] More generally, the attribute is any element detected in the scene facing the vehicle and being associated with a predefined attribute type.

[0060] A segment of the image determined by the image processing module 103 designates a portion of the image comprising at least one given attribute, for example a single given attribute. According to certain embodiments, the segment at least partially surrounds a detected contour of the attribute, i.e. at least some of the pixels of the segment are outside the detected attribute.

[0061] The image processed by the image processing module 103 and representative of the scene facing the vehicle may be the image 300 represented with reference to, and described in the following.

[0062] The image processing module 103 can further determine one or more characteristics of the image, among the following characteristics: - local characteristics such as the position, shape, size and / or type of each detected attribute, called attribute characteristics, as well as a confidence score associated with the detection of the attribute and its attribute characteristics. Indeed, in particular at night, certain attributes are poorly lit (too lit or too little lit, with poor contrast), which prevents precise and certain detection of the attributes and their attribute characteristics: in this case, the confidence score associated with the attribute and its characteristics may be low, in particular below a given threshold.The confidence score can in particular be expressed by a percentage;- an overall contrast value in the image;- an average brightness value of the image;- a local contrast value for each detected attribute, in particular in the segment comprising the detected attribute;- a local brightness value for each detected attribute, in particular in the segment comprising the detected attribute.

[0063] Thus, some of the image features are local features, specific to a given attribute, while other features are global.

[0064] Alternatively, the control module 101 is capable of determining at least some of the aforementioned characteristics on the basis of the image processed and segmented by the image processing module 103. According to another variant, the control module 101 integrates the image processing module 103. In the following, it is considered by way of illustration that the image processing module 103 and the control module 101 are distinct and that the image processing module 103 is capable of determining all of the aforementioned local and global characteristics of the image.

[0065] Such image analysis to extract both local and global characteristics is known per se. It makes it possible, as indicated in the introductory part, to precisely identify attributes of the scene facing the vehicle as well as the characteristics of these attributes during so-called “daytime” situations or when the scene is well lit, uniformly, by public lighting during so-called “nighttime” situations. Such identification may be useful for one or more driving assistance functions, implemented by a driving assistance module 104, also called ADAS, for “Advanced Driver Assistance System” in English. The ADAS module 104 is capable of implementing at least one driving assistance function. No restriction is attached to said at least one function, which depends in particular on the level of autonomy of the vehicle.Alternatively or additionally, the segmented image comprising an identification of the attributes, obtained by the image processing module 103, can be displayed on an interior screen of the vehicle, not shown in the figure, in order to assist the driver in driving the vehicle.

[0066] However, as mentioned above, the identification of certain attributes, and their characteristics, may be more uncertain during night driving situations, in particular when certain segments of the scene facing the vehicle are poorly lit: poor contrast, intensity too strong or on the contrary too weak. When an attribute is nevertheless identified, the confidence score associated with its identification, as well as with the identification of its characteristics, may be low, which may inhibit the implementation of certain ADAS functions by the ADAS module 104, or may create uncertainty for the driver when the segmented image with attribute identification is displayed on the screen.

[0067] The present invention makes it possible to improve the identification of attributes in images acquired by the camera 102. For this purpose, the control module 101 according to the invention is able to access a model, and to apply the model to at least one state of the image received from the image processing module 103, the at least one state of the image being able to comprise one or more of the following states: - a local state of an attribute identified in the image, the state being able to be determined from any combination of the local characteristics specific to the identified attribute, for example the attribute characteristics (position, type, shape, size) and the confidence score associated with the identification of the attribute;and / or- a global state of the image, the state being able to be determined from any combination of the global characteristics of the image, but also characteristics such as the respective positions of the identified attributes in the image, the local values ​​of contrast and intensity for the identified attributes, the confidence scores associated with the identifications of the attributes, or any other local characteristic, or characteristic obtained from one or more local characteristics (distance between two attributes for example).;

[0068] The number of states provided as input to the model can be predetermined. The more features considered for state determination, and the more variable their values, the greater the number of states input to the model can be.

[0069] It should be noted that the model which is the subject of the invention is distinct from the segmentation model implemented by the image processing module 103.

[0070] According to embodiments, the control module 101 is capable of determining a state for each attribute of the image.

[0071] The model accessed by the control module 101 is configured to determine an action based on a state of the image provided as input. According to the invention, as detailed in the two embodiments described below, the action makes it possible to improve the identification of attributes in the scene facing the vehicle 100. The action consists of an adaptation of the lighting or a modification of the image received from the image processing module 103.

[0072] The model may be able to determine the action:- for a local state provided as input. In this case, the action is itself local in the segment comprising the attribute for which the local state is provided as input to the model, the local action being an adaptation of the lighting in an area corresponding to the segment of the image, or a modification of the image in the segment;- for a global state provided as input. In this case, the action is global on the image and may comprise a plurality of local sub-actions on the identified attributes, the global action being an adaptation of the lighting on the entire image or in several segments of the image, or a global processing of the image, which may comprise several differentiated sub-processings in the segments determined by the image processing module 103.

[0073] In embodiments in which the model is capable of receiving a local state of an attribute as input, the model may be applied sequentially or in parallel to all local states of the respective attributes detected in the image.

[0074] No restriction is attached to the form of the model, which can be an algorithm implementing a set of rules, or “policy” in English, or an artificial neural network, such as a deep neural network for example. More generally, the model relates to any algorithm capable of determining an action on the basis of characteristics received as input. Advantageously, the model can be obtained by machine learning. According to the invention, as detailed later, the machine learning can be of the reinforcement learning type, or RL “reinforcement learning” in English. As described in the following, the reinforcement learning can be a Q-learning type learning, in which Q parameters of the model are trained, without requiring the definition of a priori rules (a “policy” in English).Alternatively, reinforcement learning can also be of the deep reinforcement learning type, particularly when the input states are determined from a large number of characteristics and / or when the model includes a large number of parameters. According to such learning, the model constructed is a deep neural network.

[0075] According to a first embodiment, the identification of attributes is improved by modifying the image previously processed by the image processing module 103. The action determined at the output of the model is therefore a modification of the image. The model is trained beforehand, so that the modified image allows a better identification of at least one of the attributes of the image by the image processing module 103, for example: higher local contrast value of the attribute and / or better confidence score associated with the identification of the attribute and the attribute characteristics. Thus, the model can be specific to the image processing (segmentation model in particular) implemented by the image detection module 103. When the image is a grayscale image, the modification can consist of modifying at least one grayscale of a pixel of the image.According to a preferred example, the modification of the image may comprise the application of a gamma transformation function fixing, for each original pixel gray level value with a pixel gray level output value. The possible actions are gamma transformation functions for different values ​​of gamut coefficient. The gamma transformation function may be applied locally to the segment of an attribute in particular, as described later in this description.

[0076] Thus, the identification of a given attribute and its attribute characteristics in the image initially received from the camera can be associated with a first confidence score, and the modification of the image by application of the model by the control module 101 allows the identification of the attribute and its attribute characteristics with a second confidence score higher than the first confidence score, and / or allows a higher local contrast value of the attribute to be obtained in the modified image than in the image initially received from the camera 102.

[0077] According to a second embodiment, the action determined by the model as a function of the state received as input, is a modification of the lighting of the vehicle 100.

[0078] In the second embodiment, the vehicle 100 advantageously comprises two front lighting modules 110, integrated respectively in a right front headlight and a left front headlight of the vehicle 100.

[0079] A single lighting module 110 is shown in the figure, to facilitate reading of the figure only. According to the second embodiment, the lighting module 110 comprises a matrix light source 111, the source comprising a set of light elements individually controllable by a control unit 112 capable of communicating with the control module 101 previously described.

[0080] The light elements of the matrix source 111 may be arranged in rows and columns, with no restrictions being attached to the number of light elements, the number of rows, or the number of columns. The control unit 112 may selectively activate / deactivate each of the light elements. For this purpose, the light elements may be electroluminescent, and the control unit selectively deactivates / activates each of the light elements of the matrix source 110 by controlling the delivered electrical power. According to certain embodiments, the control unit 112 may further control the intensity level of each light element, between at least two non-zero intensity levels, by pulse width modulation control, or PWM.

[0081] According to embodiments, the matrix light source 111 may be of the monolithic type. A so-called “monolithic” matrix light source may have a particularly high density of light elements, which makes it particularly interesting for a plurality of applications. A monolithic source involves a plurality of light-emitting semiconductor elements with submillimeter dimensions, epitaxially grown directly on a common substrate, the substrate generally being formed of silicon.Unlike conventional LED array sources, in which each elementary light element is an individually produced electronic component mounted on a substrate such as a printed circuit board (PCB), a monolithic source is considered a single electronic component, during the production of which several areas of light-emitting semiconductor junctions are generated on a common substrate, in the form of an array. This production technique makes it possible to produce light-emitting areas, each acting as an elementary light element, very close to each other. The gaps between the light elements can have submillimeter dimensions. An advantage of this production technique is the high level of pixel density that can result on a single substrate.

[0082] Thus, each light element of the matrix light source 111 is capable of projecting an elementary light beam, the combination of the elementary light beams forming a global pixelated beam, each elementary light beam forming a pixel of the global pixelated beam.

[0083] The control unit 112 is thus able to control the matrix light source 111 as a function of a photometric image transmitted by the control module 101, the photometric image defining a light intensity value and therefore a cyclic value for the PWM control of each light element. The control module 101 is able to determine the photometric image as a function of:- at least one lighting function to be performed;- as a function of at least one modification of the lighting resulting from the application of the model by the control module 101. Note that the model can be applied sequentially or in parallel to several local states relating to respective attributes detected in the image. In this case, several modifications of the lighting, which are local modifications, are taken into account in addition to the lighting function for the determination of the photometric image by the control module 101.

[0084] The modification of the lighting can thus define a modification of the light intensity of at least one light element compared to a previous photometric image. For example, the modification can consist of a modification of the light intensities of the light elements corresponding to the segment comprising an identified attribute. Indeed, the model can be trained to determine such a modification of the lighting, according to a state provided as input, descriptive of the image from the camera 102. The term “modification of the lighting” or “adaptation of the lighting” thus refers to a variation of the light intensity of at least one light element in the photometric image, compared to a previous photometric image transmitted to the lighting module 110.

[0085] The lighting modification thus includes at least one light element identifier (identified by its position in the light element matrix for example), and a variation in light intensity (downward or upward) for each identified light element.

[0086] The model according to the second embodiment is thus specific to a given camera 102 and to a given lighting module 110, and to a relative arrangement of the given camera 102 and of the given lighting module 110, in order to take into account correspondences between a pixel, or several pixels, of the image acquired by the camera 102, and a pixel, or several pixels, of the pixelated light beam formed by the lighting module 110.

[0087] The modification of the lighting resulting from the application of the model by the control module 101 according to the invention makes it possible both:- to improve the identification of attributes of the scene in a new image acquired by the camera 102 after adaptation of the lighting, and processed by the image processing module 103;- but also to improve the identification of the attribute directly by the driver. Indeed, the improvement of the confidence score associated with the attribute and the attribute characteristics implies that the adaptation of the lighting on the attribute allows a better visual identification of the attribute.

[0088] This is a diagram illustrating the steps of a current phase of a method for improving the detection of attributes in an image from a camera of a vehicle, capable of acquiring an image of a scene facing the vehicle, according to embodiments of the invention.

[0089] We distinguish the current phase, which can be implemented while the vehicle is moving, from a learning phase described later, during which the model implemented by the control module 101 is constructed, in particular trained by machine learning.

[0090] The current phase thus corresponds to a phase during which the vehicle 100 is in circulation, in particular at night, that is to say when the light intensity outside the vehicle is lower than the predetermined threshold described previously.

[0091] The current phase can be implemented in the vehicle 100 described previously with reference to the. According to certain embodiments, the control module 101 can implement all of the steps of the current phase described below.

[0092] The current phase comprises a step 200 of receiving at least one image, in color or in grayscale, from the camera 102. Note that the camera 102 can transmit images acquired at a given frequency, of several images per second in particular, for example 30 images per second.

[0093] The image can be received at step 200 by: - ​​the image processing module 103; or - the control module 101 when it integrates the image processing module 103.

[0094] The image reception in step 200 can thus be implemented at a given frequency, which depends on the frequency of acquisition of the images by the camera 102. The following steps of the current phase can be implemented for each of the images received from the camera 102. However, preferably, the following steps 201 to 204 can be applied to an image every N images received or to an image every M given seconds, N being an integer strictly greater than one.

[0095] For example, in the second embodiment, in which the lighting is modified, the following steps of the current phase can be applied to an image received every M seconds, M being between 5 and 20. Such a frequency makes it possible to update the lighting at a regular frequency so as to take into account the change in environment, while considering a certain stability of the environment over a small time step, for example of the order of one second or a few seconds.

[0096] In a step 201, the image received in step 200 is processed so as to determine at least one characteristic of the received image, and preferably at least one local characteristic of an attribute detected in the image. The analysis of step 201 may thus comprise the segmentation of the received image and the identification in at least one of the segments, of an attribute as well as one or more characteristics of associated attributes (type, size, position, and / or shape in particular). The identification of the attribute and of the at least one associated characteristic in the segment may be associated with a confidence score. The local characteristic associated with the attribute may also be a local contrast value and / or a local light intensity value in the segment comprising the identified attribute.Alternatively or additionally, the analysis of step 201 may comprise the determination of at least one global characteristic of the image, for example a global contrast value and / or a global light intensity value, as previously explained.

[0097] The image processing applied in step 201 can be implemented by:- the image processing module 103; or- the control module 101 when it integrates the image processing module 103.

[0098] Illustrates an example of an image 300 resulting from the processing of step 201. The image 300 was previously acquired by the camera in step 200, as explained previously.

[0099] Image 300 depicts a scene facing the vehicle, in a night driving situation.

[0100] In the example shown in the, the image may segment the scene into a first segment 301 in which a first attribute 311 is identified, a second segment 302 in which a second attribute 312 is identified and a third segment 303 in which a third attribute 313 is identified.

[0101] A road 304 on which vehicle 100 is traveling can also be identified.

[0102] The outlines of the attributes are clear in the, for illustrative purposes. It will be understood, however, that the outline of the attributes, as well as their identification, may be less clearly defined, particularly in night driving situations. Furthermore, in practice, the image 300 comprises grayscale or color pixels, and not clear outlines formed by black lines as shown, for simplification, in the.

[0103] Thus, errors can be made on the identification of an attribute or one of the attribute characteristics, for example on the position of the third attribute 313, or on its type (which can be identified as being an object, whereas the third attribute 313 is a pedestrian in reality). Note that an attribute can be correctly identified with a high confidence score while another attribute can be misidentified, and / or with a lower confidence score, in the same image, the brightness not being homogeneous in the scene.

[0104] In the following, it is considered that during step 201:- the first attribute 311 can be identified and associated with attribute characteristics such as its type “vehicle in the same direction of circulation”, a given position, a given size and a given shape, with a high confidence score, for example greater than 80%;- the second attribute 312 can be identified and associated with attribute characteristics such as its type “vehicle in the same direction of circulation”, a given position, a given size and a given shape, with an average confidence score, for example equal to 50%;- the third attribute 313 can be identified and associated with characteristics such as its “traffic sign” type (erroneous, compared to the real situation, the third attribute being a pedestrian), a given position (which can be offset from the real situation), a given size and a given shape, with a low confidence score, for example equal to 20%.;

[0105] Referring again to the description of the current phase of the method according to the invention, at a step 202, the control module 101 determines at least one state of the image 300 processed during step 201.

[0106] As described above, the at least one state may be:- a local state of an attribute identified in the image, the state being able to comprise any combination of the local characteristics specific to the identified attribute, for example the attribute characteristics (position, type, shape, size) and the confidence score associated with the identification of the attribute; and / or- a global state of the image, the state being able to comprise any combination of the global characteristics of the image, but also characteristics such as the respective positions of the identified attributes in the image, the local contrast and intensity values ​​for the identified attributes, the confidence scores associated with the identifications of the attributes, or any other local characteristic, or characteristic obtained from one or more local characteristics (distance between two attributes for example).

[0107] Depending on whether the model stored by the control module 101 is capable of receiving a global state or a local state as input, the control module 101 determines a global state, or one or more local states. Preferably, in the case where the model is capable of receiving a local state as input, the control module 101 determines a local state for each attribute identified in the image resulting from step 201. In the example of 1a, the control module 101 determines a first local state from local characteristics determined for the first attribute 311, a second local state from local characteristics determined for the second attribute 312 and a third local state from local characteristics determined for the third attribute 313.Alternatively, the control module 101 only determines local states for poorly identified attributes (confidence score lower than a given threshold, and / or local contrast value lower than a given threshold).

[0108] In a step 203, the control module 101 provides as input to the model, the at least one state determined in step 202, and thus obtains an action at the output of the model. When several local states are determined in step 202, the several local states are provided as input to the model sequentially, in order to obtain a series of actions at the output of the model, or the model is implemented several times in parallel for all the determined local states, in order to obtain the series of actions at the output of the model implemented several times in parallel.

[0109] At a step 204, the control module 101 implements said at least one action determined by the model.

[0110] In the first embodiment, said at least one action is a modification of the image previously processed during step 201 and originating from the camera 200. As indicated previously, the action may comprise the modification of the gray level of at least one pixel. Advantageously, each action may be the application of a gamma transformation function to a segment of an attribute in the image originating from step 200. As described later with reference to the, several different gamma transformation functions may be predetermined, the functions differing by the value of the gamma coefficient. The model may thus determine, for each local state of an attribute, a gamma coefficient value, and step 204 thus comprises the application of a gamma transformation function corresponding to the value of the determined gamma coefficient to the segment comprising the attribute.The model is thus trained to match local states with predetermined gamma coefficient values.

[0111] Examples of actions being gamma transformation functions corresponding to different values ​​of gamma coefficients are described later with reference to the.

[0112] In the example of the, applied to the first embodiment:- a first gamma coefficient can be determined by a first application of the model to the first local state determined for the first attribute 311. A first gamma transformation function defined by the first gamma coefficient is applied to the pixels of the first segment 301 comprising the first attribute 311;- a second gamma coefficient can be determined by a second application of the model to the second local state determined for the second attribute 312. A second gamma transformation function defined by the second gamma coefficient is applied to the pixels of the second segment 302 comprising the first attribute 312;- a third gamma coefficient can be determined by a third application of the model to the third local state determined for the third attribute 313.A third gamma transformation function defined by the third gamma coefficient is applied to the pixels of the second segment 302 comprising the first attribute 312.

[0113] Note that when the confidence score of an attribute is high before modification of the image 300, as is the case for the first attribute, the first application of the model can determine an action which does not transform or only slightly transforms the segment comprising the first attribute. Alternatively, the model can indicate that no action is to be taken in the first segment 301. As a further alternative, the control model 101 does not determine a local state for the first attribute.

[0114] After applying said at least one action of modifying the image 300, according to the first embodiment, the modified image can be transmitted to the image processing module 103, which can process the modified image, in order to segment the image again, and identify the attributes and the attribute characteristics, as well as a new confidence score. The model is trained in the learning phase so that the modification of the image makes it possible to improve the contrast level and / or the confidence score of at least one attribute of the image, between the image 300 processed during step 201, and a processing applied to the modified image following step 204.

[0115] The identification of the attributes in the modified image is thus improved, and descriptive information of the identified attributes can be transmitted by the image processing module 103 to the ADAS module 104 for implementing at least one driving assistance function taking as input the descriptive information of the identified attributes. In addition or as a variant, the modified image as well as the descriptive information of the identified attributes can be transmitted to a screen of the vehicle for display, which makes it possible to improve the identification of the attributes for the driver of the vehicle, and which thus reinforces the comfort and safety associated with driving the vehicle 100.

[0116] Descriptive information of the identified attributes may include, among other things, the attribute type and the confidence score.

[0117] In the second embodiment, said at least one action is a modification of the lighting of the vehicle 100. Said at least one action determined by the model thus indicates a modification of the light intensity of at least one light element of the matrix light source 111. When the action is determined for a local state of an attribute, the action is a modification of the lighting in an illuminated area corresponding to the segment of the attribute. The action can then indicate a variation in light intensity for at least one light element capable of projecting an elementary light beam in an illuminated area corresponding to the segment of the attribute. The model is trained so that the modification of the lighting in the segment improves the identification of the attribute by the driver and / or by the camera during the acquisition of a new image subsequent to the adaptation of the lighting.

[0118] Thus, not only is the identification of the attribute or attributes by the driver improved, but it is also possible to improve the identification of the attribute or attributes during the processing of step 201 applied to an image acquired by the camera 102 after the modification of the lighting.

[0119] A system for reinforcement learning a model according to embodiments of the invention is presented.

[0120] According to the principles of reinforcement learning, an agent 400 is trained to determine an action based on a state provided as input by the application of a model that it trains, the action being implemented in an environment 410. An evaluation of the action is performed by a training device 401, and a reward is determined based on the evaluation. The reward is transmitted to the agent 400 which updates parameters of the model based on the reward: the reward may in particular be positive or negative (penalty). According to reinforcement learning, the parameters of the model may be optimized to promote the obtaining of positive rewards over the long term, and not only to maximize a reward immediately obtained after an action.

[0121] The agent 400 thus optimizes parameters of a model during a preliminary learning phase, for a subsequent implementation of the model during step 201, for example in the form of an algorithm stored in the memory of the control module 101.

[0122] During training, the states provided as input to the model are determined by the training device 401 from environmental data.

[0123] The environment 410 considered (and the environment data) differs between the first embodiment and the second embodiment.

[0124] In the first embodiment, the environmental data considered are a set of training images acquired by cameras similar to the camera 102 described above, stored in a database 411. Preferably, the acquired training images are acquired by a camera of the same type as the camera 102 and having the same position in a motor vehicle, during night driving situations. Alternatively, the database may comprise a diverse set of training images acquired via different types of cameras, in night driving situations. An action from the model is a modification of the training image, and the evaluation making it possible to determine the reward linked to the action, is determined from the training image modified by the action.Alternatively, the training images are daytime images, in which the attributes can be determined with a maximum confidence score and with a high contrast level. The action evaluation can thus depend on the ability of the action to promote correct identification of the scene attributes.

[0125] In the second embodiment, the environmental data are images acquired in real time by a camera 412, of the same type as the camera 102 described previously, and arranged in an identical position in a vehicle also comprising a lighting module 413, similar to the lighting module 110 previously described. The environmental data may be acquired by more than one camera 412, respectively arranged in a fleet of vehicles comprising respective lighting modules 413.

[0126] Thus, in the second embodiment, the training is carried out in a real environment, during night driving of one or more vehicles.

[0127] In the second embodiment, the action is a modification of the lighting of the vehicle, and the training device, following the reception of an action determined by the agent 400, controls the modification of the lighting of the lighting module 413 of the vehicle having acquired the training image, which acquires an evaluation image following the modification of the lighting. The evaluation making it possible to determine the reward linked to the action is determined from the evaluation image thus received.

[0128] The steps of the learning phase implemented in the learning system described with reference to the, are described in the following with reference to the.

[0129] In a step 501, a training image is obtained by the training device 401, from the training database 411 in the first embodiment, and from the camera 412 of a vehicle in the second embodiment. In the first embodiment, when the training images are daytime images, random processing may be applied to the training images to make it more difficult to identify at least one attribute. No restriction is attached to such random processing.

[0130] In a step 502, the training device 401 processes the image received in step 501, in order to segment it and identify at least one attribute in the image, as well as at least some of the attribute characteristics described previously with reference to the current phase, as well as a confidence score for each attribute. The training device 401 can also determine one or more global characteristics of the image.

[0131] In a step 503, the training device 401 selects one or more attributes from among the identified attributes. No restriction is attached to the attribute selection criterion. For example, the attributes having the lowest confidence scores may be selected (or the attributes whose confidence scores are below a given threshold).

[0132] In a step 504, the training device 401 determines at least one state of the image. Preferably, the training device 401 determines a local state for each of the selected attributes: in this case, the model is trained to determine a local action for each selected attribute, depending on the local state. Alternatively, the training device determines a global state: in this case the model is trained to determine a global action on the image.

[0133] At a step 505, the training device 401 transmits to the agent 400:- the global state of the image;- the local states.

[0134] The agent applies the model and determines at step 506:- a global action based on the global state of the image;- local actions based on the local states of the attributes.

[0135] The at least one action determined in step 506 by the model trained by the agent 400 is implemented in the environment 410.

[0136] In the first embodiment, the global action is a modification of the training image and the local actions are modifications of the segments of the attributes of the training image. The implementation of the at least one action is therefore a modification of the training image to obtain a modified training image.

[0137] In the second embodiment, the global action is a global adaptation of the photometric image transmitted to the lighting module, and the local actions are local modifications of the photometric image, for luminous elements capable of illuminating in zones comprising the segments of the selected attributes. The implementation of the at least one action is therefore the transmission of a photometric image adapted to the lighting module of the vehicle having transmitted the training image during step 501.

[0138] In a step 508, the training device 401 evaluates the at least one action, based on predetermined evaluation metrics.

[0139] In the first embodiment, the evaluation may in particular comprise:- processing the modified training image, to identify the attributes as well as the confidence score and / or the local contrast value for each attribute;- determining for at least one attribute, a first difference between the confidence score of the attribute in the modified training image and the confidence score of the attribute obtained in step 502 in the training image (before modification) and / or a second difference between the local contrast value of the attribute in the modified training image and the local contrast value of the attribute obtained in step 502 in the training image (before modification). Preferably, the first and / or the second difference is determined for each attribute selected during step 503;- evaluating said at least one action implemented as a function of the first and second differences determined.In case of a global action determined by the model, the global evaluation can be determined based on the first and second differences for all attributes. In case of several local actions determined by the model, an evaluation of each local action can be implemented based on the first difference and / or the second difference determined for the attribute corresponding to the local action.

[0140] When the training images are initially daytime training images, in which the attributes and their attribute features are correctly identified with certainty (and which are called reference attribute features therefore), the evaluation of the at least one action may further depend on a comparison between the attribute features determined in the modified image and the reference attribute features.

[0141] In the second embodiment, the evaluation may in particular comprise:- receiving an evaluation image from the camera 412, following the modification of the lighting on the vehicle having previously transmitted the training image;- processing the evaluation image, to identify the attributes as well as the confidence score and / or the local contrast value for each attribute;- determining for at least one attribute, a first difference between the confidence score of the attribute in the modified training image and the confidence score of the attribute obtained in step 502 in the training image (before modification) and / or a second difference between the local contrast value of the attribute in the modified training image and the local contrast value of the attribute obtained in step 502 in the training image (before modification).Preferably, the first and / or the second difference is determined for each attribute selected during step 503; - the evaluation of said at least one action implemented as a function of the first and second differences determined. In the case of a global action determined by the model, the global evaluation can be determined as a function of the first and second differences for all the selected attributes. In the case of several local actions determined by the model, an evaluation of each local action can be implemented as a function of the first difference and / or the second difference determined for the attribute corresponding to the local action.

[0142] In a step 509, the training device 401 determines at least one positive or negative reward for each evaluation of an action. In the case of a global action determined by the model, a single positive or negative reward is determined. In the case of several local actions determined by the model, a global reward can be determined based on the evaluations determined for the local actions, or, alternatively, a reward can be determined for each evaluation determined for a local action.

[0143] Said at least one reward is transmitted by the training device 401 to the agent 400 at a step 510 for an update of parameters of the model according to said at least one reward. When several rewards for local actions are transmitted, the agent 400 updates the parameters according to each reward corresponding to each local action determined during the step 506 previously described.

[0144] The method can then return:- to step 504 by determining at least one state for the modified training image in the first embodiment or for the evaluation image in the second embodiment, by considering the modified training image or the evaluation image as a training image for the new iteration of the following steps. The at least one state is then submitted to the model in step 505 and the following steps are implemented again. A return to step 504 can be implemented when a performance criterion in the modified training image or in the evaluation image is not met.The performance criterion may for example be a minimum confidence score threshold for each attribute or a minimum local contrast value threshold for each attribute; or- in step 501, to repeat the following steps from a new training image acquired in the environment 410, either a new training image from the database 411 in the first embodiment or a new image acquired by one of the cameras 412 (for example for a vehicle other than the one having provided the training image during the previous iteration) in the second embodiment. A return to step 501 may be implemented when the performance criterion in the modified training image or in the evaluation image is reached.

[0145] Thus, such reinforcement learning allows building a model that ensures improvement of the detection of the attribute in the modified image or in a new image acquired after adaptation of the lighting. Indeed, the rewards depend on the improvement or not of the confidence score and / or the local contrast value of each attribute (first and second differences of the evaluation).

[0146] The present examples of actions for modifying an image or segments of an image, determined by the model according to the first embodiment of the invention.

[0147] As indicated previously, each action according to the first embodiment is a modification of the image (during step 506 during the learning phase and during step 203 during the current phase), which can be a global modification of the image or a modification of a segment comprising a detected attribute.

[0148] Changing pixels in an image, for example pixels in a segment containing a detected attribute, can be a gamma transformation function, defined by a gamma parameter.

[0149] The model can thus determine a value of the gamma parameter from a set of predetermined values ​​of the gamma parameter, and the determined gamma parameter value defines the output action of the model, to be applied to a segment of the image or to the entire image.

[0150] Shows examples of gamma transformation functions for five distinct values ​​of gamma parameter.

[0151] The five curves 601 to 603 are representative of a gamma transformation function corresponding to a given value of gamma parameter.

[0152] Each curve maps an original grayscale value, on the abscissa, to an output grayscale value, on the ordinate. This makes it possible to modify a set of grayscale pixels. When a color image is received at step 200 of the current phase, the color image can be converted to a grayscale image before applying the following steps of the current phase.

[0153] A first modification action corresponds to the first curve 601, according to which the gray level of the pixels of the image or segment are generally decreased, and the contrast is reduced. In contrast, the fifth action 605 increases the gray levels of the pixels of the image, regardless of the original gray level value in the image. Other transformations such as the second and third actions corresponding to the second and third curves 602 and 603 increase the gray level of some pixels and decrease the gray level of some pixels, depending on the original gray level value in the image. The fourth curve 604 increases the gray level value but with a higher contrast than for the fifth curve 605.

[0154] The five gamma parameter values ​​corresponding to the five curves 601 to 605 are given for illustrative purposes, and more than five gamma parameter values ​​(and in any case at least two distinct values) may be predefined, so that the model according to the first embodiment selects one of the gamma parameter values ​​as an output action of the model, upon receipt of a local or global state of the image.

[0155] The present invention shows the structure of the control module 101 according to embodiments of the invention.

[0156] The control module 101 comprises a processor 701 configured to communicate unidirectionally or bidirectionally, via one or more buses or via a direct wired connection, with a memory 702 such as a “Random Access Memory” type memory, RAM, or a “Read Only Memory” type memory, ROM, or any other type of memory (Flash, EEPROM, etc.). Alternatively, the memory 702 comprises several memories of the aforementioned types.

[0157] The memory 702 is capable of storing, permanently or temporarily, at least some of the data used and / or resulting from the implementation of the steps of the method according to the invention illustrated with reference to the.

[0158] In particular, the memory 702 may be capable of storing the model trained during the learning phase described with reference to the.

[0159] The processor 701 is able to execute instructions, stored in the memory 702, for the implementation of steps 202 to 204 of the current phase, described with reference to the. In certain embodiments, the image processing module 103 can be integrated in the control module 101, and in this case the processor is able to execute instructions, stored in the memory 702, for the implementation of steps 200 to 204 of the current phase. Alternatively, the processor 701 can be replaced by a microcontroller designed and configured to carry out steps 202 to 204 of the current phase, described with reference to the, or to carry out steps 200 to 204 of the current phase, in certain embodiments.

[0160] The control module 101 comprises a first interface 703 capable of communicating with the image processing module 103, in order to receive the image processed during step 201, and in order to transmit the modified image in the first embodiment.

[0161] In the embodiments in which the control module 101 integrates the image processing module 103, the first interface 703 is capable of receiving images from the camera 102 of the vehicle 100.

[0162] The image generation device 101 may comprise a second interface 704, in the second embodiment, in order to transmit a photometric image to the lighting module 110 during the step 204 previously described, the photometric image being adapted from the at least one lighting adaptation action determined in step 203.

[0163] The present invention is not limited to the embodiments described above as examples; it extends to other variants.

Claims

A method for detecting an attribute (311-313) in an image (300) acquired by an image acquisition device (102) of a vehicle (100), the image acquisition device being arranged to acquire images representative of a scene facing the vehicle, the method comprising the following steps of a current phase:- obtaining (201) at least one characteristic of at least one attribute detected in the image and included in a segment of the image;- determining (202) at least one state of the image as a function of the at least one characteristic obtained;- determining (203), for each determined state, an action as a function of the determined state, by applying a model designed and parameterized to receive as input an image state and to determine as output an action dependent on the state received as input;- implementing (204) said action, in which the action is a modification of the image to obtain a modified image, or an adaptation of the vehicle's lighting. Method according to claim 1, wherein the at least one characteristic obtained comprises at least one confidence score associated with the detection of the at least one attribute (311-313) and / or a local contrast value in the segment (301-303) comprising the at least one attribute. Method according to claim 2, wherein the model is parameterized so that, for at least one given attribute detected in the image, a confidence score associated with the detection of the given attribute and / or a local contrast value of the at least one given attribute, is higher in the modified image or in a new image acquired by the camera after adaptation of the lighting of the vehicle, than in the image (300). The method of claim 2 or 3, wherein the at least one obtained characteristic further comprises one or more of the following characteristics:- one or more attribute characteristics among an attribute type, an attribute position, an attribute shape and an attribute size;- a local brightness value in the segment of each detected attribute;- an overall contrast value in the image;- an average brightness value in the image. Method according to one of the preceding claims, further comprising a learning phase, the learning phase (501-510) comprising reinforcement learning of the model, from training images. Method according to claim 3 and claim 5, wherein the reinforcement learning comprises an evaluation (508) of each action resulting from the model from a state determined for a training image, the evaluation being based, for a given action, on:- a difference between a first initial confidence score associated with an attribute detected in the training image, and a second confidence score determined for the attribute after implementation of the action, in the modified training image or in an evaluation image received following the lighting adaptation;- a difference between a local contrast value determined for a segment comprising the attribute detected in the training image and a local contrast value determined for the segment comprising the attribute after application of the action;wherein, for each evaluation, a reward is determined (509), said model being updated (510) according to said reward. Method according to claim 5 or 6, wherein the reinforcement learning is based on a Q-learning type algorithm. Method according to one of the preceding claims, wherein the determination (202) of the at least one state of the image (300) comprises the determination of a local state for each attribute (311-313) detected in the image, as a function of at least one characteristic determined for said attribute, and in which the model is applied (203) to each local state to obtain a local action, the implementation (204) of each local action determined for an attribute comprising a modification of the segment of the image comprising the attribute to obtain the modified image or an adaptation of the lighting in an illuminated area corresponding to the attribute. A method according to one of the preceding claims, wherein each attribute (311-313) detected in the image (300) is among one of the types comprising a pedestrian, a road infrastructure element, an object or other vehicle. Method according to one of claims 1 to 9, wherein the action is a modification of the image (300) to obtain a modified image, the method further comprising the detection of said at least one attribute in the modified image and the transmission of the modified image to a driving assistance module (104) of the vehicle (100) capable of implementing at least one driving assistance function as a function of the at least one attribute detected in the modified image, and / or the transmission of the modified image and descriptive information of the at least one attribute detected in the modified image to a screen of the vehicle. The method of claim 10, wherein the image modification is a gamma transformation function defined by a gamma parameter value selected from a predetermined set of gamma parameter values, each gamma transformation function mapping an input gray level to an output gray level. Method according to one of claims 1 to 9, in which the action is an adaptation of the lighting of the vehicle, the vehicle comprising a lighting module (110) comprising a matrix light source (111), the matrix light source comprising a set of individually controllable light elements, and in which the adaptation of the lighting comprises a modification of a light intensity value of at least one light element capable of emitting an elementary light beam towards an illuminated zone corresponding to at least one segment comprising an attribute (311-313) detected in the image. Computer program comprising instructions for implementing the method according to one of the preceding claims, when these instructions are executed by a processor (701). Control module (101) of a motor vehicle (100), the control module being capable of obtaining at least one characteristic of at least one attribute (311-313) detected in an image (300) and included in a segment (301-313) of the image, the image coming from an image acquisition device (102) of the vehicle, the image acquisition device being arranged to acquire images representative of a scene facing the vehicle, the control module comprising a processor (702) configured to:- determine at least one state of the image as a function of the at least one characteristic obtained;- determine, for each determined state, an action as a function of the determined state, by applying a model designed and parameterized to receive as input an image state and to determine as output an action dependent on the state received as input;- implementing said action, wherein the action is a modification of the image to obtain a modified image, or an adaptation of the lighting of the vehicle.; Motor vehicle (100) comprising an image acquisition device (102) arranged to acquire images representative of a scene facing the vehicle, and a control module (101) according to claim 14. Motor vehicle (100) according to claim 15, wherein the action is a modification of the image (300) to obtain a modified image, wherein the vehicle comprises an image processing module (103) capable of detecting said at least one attribute (311-313) in the image and in the modified image, the vehicle further comprising a driving assistance module (104) capable of implementing at least one driving assistance function as a function of the at least one attribute detected in the modified image, and / or a screen of the vehicle capable of receiving the modified image and descriptive information of the at least one attribute detected in the modified image, for display on said screen. Motor vehicle (100) according to claim 15, wherein the action is an adaptation of the lighting of the vehicle, the vehicle comprising a lighting module (110) comprising a matrix light source (111), the matrix light source comprising a set of individually controllable light elements, and wherein the adaptation of the lighting comprises a modification of a light intensity value of at least one light element capable of emitting an elementary light beam towards an illuminated area corresponding to at least one segment (301-303) comprising an attribute (311-313) detected in the image.

Images