Training a neural network to assist in driving a vehicle by identifying hard-to-observe boundaries

Abstract

The present invention relates to a training method for a neural network, wherein the neural network needs to detect the boundaries of a traffic lane of an autonomous driving vehicle, and the training method comprises: a first step (10-20), wherein a first image and a second image of a section of the traffic lane are acquired to record the first image and the second image, a second step (30-70), wherein an aerial image of the traffic lane is constructed by the second image, and then the boundaries are detected from the aerial image, and the elements are detected from the first image, and then, when these elements extend these boundaries, these elements are recorded in a database, and when these elements do not extend these boundaries, each element is recorded in the database corresponding to the boundaries, and a third step (80), wherein the neural network is trained by the records in the database.

Description

Technical Field

[0001] The present invention relates to at least partially automatically (or autonomously) driven and optionally motorised land vehicles and more precisely to the training of neural networks for detecting the limits of a traffic lane in which such a vehicle is travelling. Background Art

[0002] In the following, a land vehicle is considered to be at least partially automatically (or autonomously) driven when it is driven on a highway automatically (partially or completely (without driver intervention)) during an automatic driving phase or when it is driven on a highway manually (and therefore with driver intervention on the steering wheel and / or pedals) during a manual driving phase. For example, the automatic (or autonomous) driving of an (at least partially automatically driven) vehicle may be intended to partially or completely steer the vehicle or to provide any type of assistance to a natural person driving the vehicle. The vehicle thus covers any automatic (or autonomous) driving from level 1 to level 5 in the classification of the OICA (International Organization of Motor Vehicle Manufacturers).

[0003] Furthermore, a “land vehicle” is understood here to be any type of vehicle that can be driven on roads, in particular a motor vehicle, a utility vehicle, a moped, a motorcycle, a minivan, a bus, a storage robot in a warehouse or a road car.

[0004] There are numerous driving assistance methods for assisting the driving of (at least partially autonomous) vehicles, in particular driving assistance methods which aim to control the driving based on information determined from the vehicle environment, which is based on an analysis of data (in particular digital image data) acquired from this vehicle environment by sensors.

[0005] Some of this environmental information is the limits of the traffic lane in which the vehicle is traveling, based on which it is possible to determine in particular the number of traffic lanes on the same road, the traffic lane in which the vehicle is traveling, the relative position of the vehicle on the traffic lanes and the maneuvers that the vehicle has to perform (e.g. by changing the traffic lane to overtake or to remain in its traffic lane). It is therefore necessary to correctly determine at least the limits of the traffic lane taken by the vehicle (in particular relative to a reference system associated with the vehicle).

[0006] However, it often happens today that the delimitation is temporarily confused with elements present on the traffic lane, such as reflections or asphalt seams, not only due to the too mediocre resolution of the surrounding images acquired by the onboard cameras and due to the positioning of these onboard cameras, but also due to changing meteorological and / or lighting conditions. In fact, it is understood that rain, ice, fog or near-ground lighting can cause changes in the acquired color of the delimitation, which can deceive the algorithm and thus lead to erroneous driving decisions. Moreover, it can happen that the delimitation is partially obscured and is therefore not visible on the acquired surrounding image, especially in the case of snow, dense fog or mud (or any other substance).

[0007] Indeed, a method of assistance using a neural network previously trained with the aid of a database for determining limits, the database being constructed on the basis of environmental data acquired by at least one sensor of the analysis vehicle during a driving phase, and a filtering module responsible for filtering these determined limits according to the confidence parameter values ​​assigned to them, so as to retain only those determined limits assessed to be sufficiently robust (or reliable). However, this filtering scheme proves to be "curative" rather than preventive, since it merely picks out bad detections resulting from imperfect acquisition and / or analysis, but does not correct said acquisition and said analysis (so as to be able to carry out this same picking before the vehicle enters an area comprising the limits detected upstream of the vehicle, and thus allowing these acquisitions and analyses to be taken into account in the vehicle's analysis and decision process). Therefore, this filtering scheme does not guarantee that the filtered limits are effectively used to control the autonomous driving of the vehicle.

[0008] The object of the present invention is therefore in particular to improve the described situation. Summary of the invention

[0009] To this end, the invention provides, in particular, a method for training a neural network, which itself is used to detect the limits of a traffic lane in which a vehicle is traveling at least partially autonomously.

[0010] The method is characterized in that the method comprises:

[0011] A first step, in which first and second images of different first and second sections of a traffic lane are acquired respectively by means of a first sensor and a second sensor respectively mounted in different orientations on an analysis vehicle, the analysis vehicle simulating the (autonomous) vehicle traveling on the traffic lane, and each first image and each second image is recorded,

[0012] a second step, in which a so-called aerial image of the traffic lane is constructed in a top view by means of the recorded successive second images, the limits of the traffic lane are then determined from the aerial image, and elements capable of characterizing the limits of the traffic lane are determined from each first image, it is then determined whether these determined elements extend these determined limits, and in the affirmative case, each determined element is recorded in a database, and in the negative case, each determined element is recorded in the database in correspondence with the determined limits, and

[0013] The third step is to train the neural network with the records of the database so that the neural network learns that when there is an element just determined from the first image, the neural network should either use the element when it is not stored corresponding to the boundary, or use the boundary stored corresponding to the element.

[0014] This consideration of the vehicle context allows a significantly more complete and more precise training (or learning) of the neural network, which enables the neural network to determine limits more easily, more accurately and more reliably during the vehicle's driving assistance phase and therefore has a reduced detection error rate (in particular a "false positive" rate) relative to prior art neural networks.

[0015] The training method according to the invention may include the following further features which may be used alone or in combination:

[0016] In a first step of the training method, a first image of successive first sections of the traffic lane and a second image of successive second sections of the traffic lane may be acquired, each of the successive first sections being located at a predefined distance in front of the analysis vehicle and each of the successive second sections being located in the immediate vicinity of the analysis vehicle;

[0017] In a first step of the training method, for example, a second image of a second successive section of the traffic lane can be acquired, the second successive section being located in the immediate vicinity of the rear of the analysis vehicle;

[0018] In the first step of the training method, each first image and each second image may be recorded corresponding to at least one background information. In this case, in the second step of the training method, when the determined element extends the determined boundary, each determined element may be recorded in the database corresponding to each recorded background information corresponding to each determined element, and in the negative case, each determined element may be recorded in the database corresponding to the determined boundary and to each recorded background information corresponding to each determined element, and in the third step, the neural network may be trained by the records of the database so that the neural network learns that when there is at least one current background information and there is an element just determined from the first image, the neural network needs to either use the element when it is stored corresponding to the current background information or use the boundary stored corresponding to the element with the current background information;

[0019] In the first step of the training method, at least one background information selected from the following information may be used, for example: meteorological information related to the current environment of the analysis vehicle, light information characterizing the brightness of the exterior in front of the analysis vehicle, material shading information received in the analysis vehicle via a wave channel, activation of an exterior lighting system of the analysis vehicle, and the angular position of the sun relative to the horizon located in front of the analysis vehicle;

[0020] In the first step of the training method, the meteorological information may be selected from rain, ice, snow and fog;

[0021] In the first step of the training method, the meteorological information may be received into the analysis vehicle via a radio channel, or provided by a detector mounted on the analysis vehicle;

[0022] In a first step of the training method, the angular position of the sun can be determined from acquired images or can be determined based on the current time and the current geographical location of the analysis vehicle.

[0023] The present invention also provides a computer program product comprising a set of instructions capable of implementing a training method of the type described above when executed by a processing unit to train a neural network for detecting the limits of a traffic lane in which a vehicle is traveling at least partially autonomously.

[0024] The present invention also provides a training device for training a neural network, wherein the neural network itself is used to detect the boundaries of a traffic lane in which a vehicle that is at least partially autonomously driving is traveling.

[0025] The training device is characterized in that the training device includes at least one processor and at least one memory, and the at least one processor and the at least one memory are configured to perform the following operations:

[0026] obtaining first and second images, respectively, of different first and second sections of a traffic lane, the first and second images being acquired by first and second sensors, respectively, mounted along different orientations on an analysis vehicle, the analysis vehicle simulating the (autonomous) vehicle traveling on the traffic lane, and then

[0027] A so-called aerial image of the traffic lane is constructed in a bird's-eye view by means of the recorded successive second images.

[0028] Determine the limit of the traffic lane from the aerial image, and determine from each first image an element capable of characterizing the limit of the traffic lane, and then

[0029] determining whether the determined elements extend the determined limits and, in the affirmative, recording each determined element into a database, or, in the negative, recording each determined element into the database corresponding to the determined limits, and

[0030] The neural network is trained by the records of the database so that it learns that, when an element just determined from the first image is present, it must either use the element if it is not stored corresponding to a boundary, or use the boundary stored corresponding to the element.

[0031] The present invention also provides a computer, which includes, on the one hand, a neural network to be trained, wherein the neural network is used to detect the boundaries of a lane in which a vehicle that is at least partially autonomously driving is traveling, and, on the other hand, includes a training device of the type described above, wherein the training device is connected to the neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Other features and advantages of the present invention will become more apparent from the following detailed description and the accompanying drawings, in which:

[0033] Figure 1 Schematically and functionally shows an embodiment of a computer equipped with an embodiment of a training device according to the invention and with a neural network to be formed, and

[0034] Figure 2 An example of an algorithm implementing the training method according to the invention is schematically shown. DETAILED DESCRIPTION

[0035] The invention has for its object in particular to provide a training method and an associated training device DE, which are used to allow the training of a neural network RN, which itself is used to subsequently detect the limits of a traffic lane in which a vehicle is traveling at least partially autonomously, including when these limits are difficult to observe.

[0036] In the following, as a non-limiting example, it is considered that the automatically (or autonomously) driven vehicle is of the motor type. The vehicle is, for example, a car. However, the invention is not limited to this type of vehicle. The invention relates in fact to any type of land vehicle that is at least partially automatically driven and that is capable of traveling on a land lane. Thus, the vehicle may also be, for example, a utility vehicle, a motorbike, a motorcycle, a minivan, a coach, a storage robot in a warehouse or a road car.

[0037] As described above, the invention provides in particular a training method for allowing the training of a neural network RN, which itself is then used to detect the limits of a traffic lane in which a vehicle is traveling at least partially autonomously. It is understood that this neural network RN is intended to be equipped, when trained, on the above-mentioned vehicle (hereinafter so-called autonomous vehicle). More precisely, this neural network RN is intended to be a component of a driving assistance device equipped on an autonomous vehicle in order to at least determine the limits of the traffic lane in which the vehicle is traveling.

[0038] The training method according to the invention can be implemented at least partially by a training device DE, which for this purpose comprises at least one processor PR, such as a digital signal processor (or DSP (“Digital Signal Processor”)) and a memory MD, such as Figure 1 In addition, the implementation of the training method requires at least one analysis vehicle, which simulates the automatic vehicle and is equipped with at least a first sensor and a second sensor installed along different first orientations and second orientations, respectively, so as to respectively obtain first images and second images of different first sections and second sections of the traffic lane on which the analysis vehicle travels.

[0039] For example, each of these first and second sensors may include at least one camera responsible for acquiring digital images from the environment of the analysis vehicle. As described further below, these sensors are used to acquire environmental data that can determine elements of the traffic lane on which the analysis vehicle is traveling.

[0040] As a non-limiting example, the first sensor may acquire a first image of a first successive segment of the traffic lane, each of which is located at a predefined distance in front of the analysis vehicle (for example between 0 m and 100 m (upstream of the analysis vehicle)), and the second sensor may acquire a second image of a second successive segment of the traffic lane, each of which is located in the immediate vicinity of the analysis vehicle (for example between 0 m and 5 m (upstream and / or downstream of the analysis vehicle)). To this end, the first sensor may, for example, comprise at least one camera responsible for acquiring digital images at least from in front of the analysis vehicle (from a "long distance"), the at least one camera being mounted in the front of the analysis vehicle (for example on the windshield or on the interior rearview mirror or on the roof or in a pillar or in a radiator grille (in the context of a sufficiently high vehicle such as a pickup truck or a bus)), and optionally on at least one of the two sides of the analysis vehicle. Such a first sensor is typically what a person skilled in the art calls a "frontal" sensor, since it has a main viewing axis which is slightly inclined relative to the traffic lane.

[0041] As for the second sensor, it can, for example, acquire a second image of a second, successive section of the traffic lane, which is located in the immediate vicinity of the rear of the analysis vehicle. Such a second sensor is usually what a person skilled in the art calls a "body" sensor, since it usually comprises at least one camera fixed to the vehicle body and in the lower part, which has a main viewing axis which is heavily inclined relative to the traffic lane (so as to give an almost top-down image). This type of sensor is responsible, for example, for providing an image of the rear surroundings for parking assistance (hence the name reversing camera when arranged at the rear of the vehicle). However, the second sensor can, for example, acquire a second image of a second, successive section of the traffic lane, which is located in the immediate vicinity of the front of the analysis vehicle, in order to construct an aerial view by direct observation or by memorizing (and juxtaposing) continuously acquired quantities in order to extend the surface thus observed "from above".

[0042] like Figure 1As shown above without limitation, the training device DE can be at least partially a component of a computer OS, which also includes the neural network RN to be trained (or formed). However, this is not necessary. In fact, the training device DE can include an inherent computer and can be connected to the computer OS including the neural network RN to be trained (or formed). Therefore, the training device DE can be implemented in the form of a combination of electrical or electronic circuits or devices (or "hardware") and software modules (or "software"). The memory MD is a random access memory in order to store instructions, which are used to implement at least part of the training method through the processor PR.

[0043] Note that the computer OS may be fixed or portable or may be part of a server. The computer may be mounted in the analysis vehicle or may be moved out for large-scale processing (optionally in the cloud).

[0044] The training method according to the present invention comprises the following reference Figure 2 The first steps 10-20, the second steps 30-70 and the third step 80 are described.

[0045] In a sub-step 10 of the first step 10-20 of the training method, first images i1 and second images i2 of different first and second sections of the traffic lane of the analysis vehicle are acquired by a first sensor and a second sensor (installed along different orientations) of the analysis vehicle, respectively. Then, in a sub-step 20 of the first step 10-20, each first image i1 and each second image i2 are recorded.

[0046] It is understood that the images i1 and i2 are acquired during a driving phase of the analysis vehicle, these images being recorded in a memory thanks to a computer on board the analysis vehicle.

[0047] In sub-step 30 of the second step 30-70 of the training method, a so-called aerial image of the traffic lane in question is constructed (by the processor PR and the memory MD) from the successive second images i2 (recorded in the first steps 10-20) in a top view. For example, between 3 and 5 successive second images i2 are used to construct the aerial image ia. It is also possible to think in terms of distance rather than time (or number of images), especially when the vehicle is traveling at a low speed. In this case, the map may not be updated in order to save computing resources and / or to avoid losing old data.

[0048] Then, in sub-step 40 of the second step 30-70 of the training method, the limits dj of the traffic lane are determined from the aerial image ia (constructed in sub-step 30) (by the processor PR and the memory MD), and elements ej capable of characterizing the limits dj of the traffic lane are determined from each first image i1. It is important to note that the limits dj determined from the aerial image ia are regarded as the limits of the traffic lane in a positive manner, while the elements ej may be these same limits dj, or may be reflections or asphalt seams.

[0049] Then, in a sub-step 50 of the second step 30-70 of the training method, it is determined (by the processor PR and the memory MD) whether these determined elements ej extend these determined limits dj. To this end, it is possible to determine, for example, by calculation, whether the geographical position of the element ej corresponds to the geographical position of the corresponding determined limit dj (optionally by taking into account its own form, whether it corresponds to an extended extrapolation of the determined limit (dj)), and whether the form of the element ej corresponds to the form of the determined limit dj (optionally by taking into account its own form, whether it corresponds to an extended extrapolation of the determined limit (dj)).

[0050] In the affirmative case (and therefore in the case of a valid extension (or correspondence)), in a sub-step 60 of the second step 30 - 70 of the training method, each determined element ej is recorded (processor PR and memory MD) into a database BD.

[0051] Conversely, in the negative case (and therefore in the absence of a valid extension (or correspondence)), in a sub-step 70 of the second step 30-70 of the training method, each determined element ej is recorded in the database BD corresponding to the determined limit dj (processor PR and memory MD). The database BD thus stores a first record (ej) and a second record (ej', dj').

[0052] like Figure 1 As indicated above without limitation, the database BD can be a component of the computer OS so as to be directly accessible by the training device DE and the neural network RN. However, the database (BD) can be a component of another computer (optionally a remote computer) or a server (which is connected to the computer OS and can be accessed via at least one communication network).

[0053] Then, in a sub-step 70 of a third step 80, the neural network RN is trained (processor PR and memory MD) by means of the first record and the second record of the database BD. This type of training serves to make the neural network RN learn that, when there is an element just determined from the first image i1 (by the first sensor of the future automatic vehicle), the neural network must either use this element (i.e. the first record) if it is not stored corresponding to a limit, or use the limit stored corresponding to this element (in the second record).

[0054] It is noted that this type of training (the so-called "supervised" type of training) is common and well known to those skilled in the art.

[0055] It is understood that during training, the neural network RN receives as input at least one first image i1 from which it determines an element ej capable of characterizing a limit dj. The neural network RN then searches the database BD for a first record or a second record containing this element ej. When the neural network finds the first record, the neural network RN learns that the same element ej detected in the future is effectively the limit dj of the lane of passage of the automated vehicle. Conversely, when the neural network finds the second record, the neural network RN learns that the same element ej detected in the future is not the limit of the lane of passage of the automated vehicle and that, instead, the neural network has to use the limit dj initially associated with this element ej.

[0056] When the neural network RN is shown enough examples (that is, pairs of [input quantity (image), associated output quantity (limit)]), the neural network learns an inherent (appropriate) representation of the limits in the image. Therefore, it is necessary to build a database with annotations (that is, indicating the output quantity to be obtained for any input data) that is sufficiently "rich" (in terms of the coverage of the encountered situations) to make the neural network RN more efficient in more situations. The provided solution makes it possible to obtain these annotations through filtering based on the analysis of detections (common on the frontal image and on the local aerial view), and thus to automatically build the associated learning database. This makes it possible to handle more situations, and therefore to cover more situations in the learning database, and therefore to make the neural network RN (and the decision system placed in the neural network) more robust (especially for restricted situations) and more efficient (especially with fewer false positives).

[0057] The training (or learning) of the neural network RN is significantly more complete and more precise, which enables the neural network, when it is part of a driving assistance device installed on an automated vehicle, to determine limits more easily, more precisely and more reliably during the driving assistance phase of the automated vehicle.

[0058] It is noted that in the sub-step 20 of the first step 10-20, each first image i1 and each second image i2 is particularly advantageously recorded corresponding to at least one context information ic. Thus, tuples of the type (i1, ic) and (i2, ic) are obtained. It is understood that for this purpose, the analysis vehicle is directly supplied with context information ic, automatically or with assistance or monitored by a technician. It is noted that the training device DE can optionally include an additional module present in the on-board computer, which is responsible for controlling the recording of the tuples (i1, ic) and (i2, ic).

[0059] In this case, in sub-step 60 of the second step 30-70 of the training method, when the determined element extends (or corresponds to) the determined boundary (positive case), (processor PR and memory MD) record each determined element ej in the database BD corresponding to each recorded background information ic corresponding to each determined element (that is, as a component of the multi-tuple (i1, ic) containing the first image i1, in which each determined element was detected).

[0060] Conversely, in the negative case (and therefore in the absence of a valid extension (or correspondence)), in a sub-step 70 of the second step 30-70 of the training method, each determined element ej is recorded into the database BD (by the processor PR and the memory MD) in correspondence with the determined delimitation dj and with each recorded context information ic corresponding to each determined element (that is, as part of the tuple (i1, ic) containing the first image i1 in which each determined element was detected). The database BD thus stores first type records (ej, ic) and second type records (ej', dj', ic').

[0061] Next, in a sub-step 70 of a third step 80, the neural network RN is trained (by the processor PR and the memory MD) by means of the first record and the second record of the database BD. Thus, the neural network RN learns that when there is at least one current context information (relating to the future automated vehicle) and there is an element just determined from the first image i1 (by the first sensor of the future automated vehicle), the neural network must either use the element (in the first record) when it is stored corresponding to the current context information or use the limit (in the second record) stored corresponding to the current context information and to the element.

[0062] In the presence of this optional configuration (use of context information), during training, the neural network RN receives as input on the one hand a first image i1 from which the neural network determines an element ej capable of characterizing a limit dj and on the other hand at least one context information ic. The neural network RN then searches the database BD for a first record or a second record comprising the element ej and the received context information ic. When the neural network finds the first record, the neural network RN learns that the same element ej detected in the future under the same context conditions (ic) is effectively the limit dj of the lane of travel of the automated vehicle. Conversely, when the neural network finds the second record, the neural network RN learns that the same element ej detected in the future under the same context conditions (ic) is not the limit of the lane of travel of the automated vehicle and that, instead, the neural network has to use the limit dj initially associated with the element ej under the same context conditions (ic).

[0063] For example, in sub-step 20 of the first step 10-20, (processor PR and memory MD) may use at least one background information selected from: meteorological information related to the current environment of the analysis vehicle, light information characterizing the external brightness in front of the analysis vehicle, shading information of materials (such as mud, earth or sand) received in the analysis vehicle through a channel, activation of the external lighting system of the analysis vehicle, and the angular position of the sun relative to the horizon located in front of the analysis vehicle.

[0064] As an example, each weather information may be selected from rain, ice, snow, and fog.

[0065] As another example, each weather information can be received by the analysis vehicle via a channel (through the communication module of the analysis vehicle, or a smart phone (or "smartphone") or a vehicle-mounted tablet carried by a technician), or can be provided by a detector (such as a rain detector) installed on the analysis vehicle.

[0066] Also as an example, the angular position of the sun can be determined from the acquired images (in particular the first image i1) or (when the navigation aid is part of a portable device or a smart mobile phone or a tablet carried by a technician) based on the current moment and the current geographic location of the analysis vehicle (the current geographic location is provided, for example, by a navigation aid permanently or temporarily included in the analysis vehicle).

[0067] Note that, when the neural network RN is part of a driving assistance device equipped on board an automatic vehicle, when said neural network has been trained by taking into account (one or more) pieces of background information, said neural network is informed of each piece of background information to be taken into account (optionally solely based on the training it has received) by a computer of the automatic vehicle or by a communication module permanently or temporarily present in the automatic vehicle (and as part of, for example, a smart mobile phone or a passenger tablet).

[0068] It is also noted that, when the neural network RN is trained and is part of a driving assistance device, it is no longer necessary a priori to connect to a database similar to the training database BD. However, it is conceivable that the driving assistance device includes a database that can be supplemented by learning of new situations during the driving phases of the automatic vehicle. In this case, it is also conceivable that the neural network RN continues to learn and thus reduces the error rate of limit detections over time due to these supplements to the database. However, it is currently preferred to build a learning database for new situations (situations in which the system performs poorly, which can be detected by a provided solution that is the same as the solution that can filter the good detections to supply the initial learning database), and then let the neural network RN learn through these new situations more than the initial ones. This last step can be carried out under the control of the manufacturer (for example, by obtaining these situations from the cloud (or "cloud") of all vehicles, for example at the time of passages in concession), and then redistribute the updated neural network by means of an approved update. An intermediate approach aims at implementing updates of local (that is to say, at each vehicle position) but frequently travelled segments, in which case a "correction" of the filter portion can be confirmed by the number of occurrences.

[0069] It is also noted that the present invention also provides a computer program product comprising a set of instructions, which, when executed by a processing component of the electronic circuit (or hardware) type (such as a processor PR), is capable of implementing the training method for training a neural network RN as described above, wherein the neural network is used to detect the boundaries of a traffic lane in which a vehicle traveling at least partially autonomously is traveling.

[0070] Also note that in Figure 1, a training device DE with a random access memory MD and a processor PR is shown very schematically, the processor may comprise an integrated (or printed) circuit or a plurality of (or printed) integrated circuits connected by wired or wireless connections. An integrated (or printed) circuit is understood to be any type of device capable of performing at least one electrical or electronic operation. However, as Figure 1 As shown above, without limitation, the training device DE may also include a large-capacity memory MM, which is used in particular for the temporary storage of acquired images, the temporary storage of optional background information, the temporary storage of geographical locations and the temporary storage of intermediate data that are optionally involved in all calculations and processing. In addition, the training device DE may also include an input interface IE, which is used to receive at least acquired images, optional background information and geographical locations, in order to use these information in calculations or processing, optionally after collating and / or demodulating and / or amplifying them in a manner known per se by means of a digital signal processor PR'. In addition, the training device DE may also include an output interface IS, which is used in particular for transmitting commands and instructions to the neural network RN during training.

[0071] It is also noted that one or more steps and / or one or more sub-steps of the training method can be performed by different components. Thus, the training method can be implemented by multiple digital signal processors, random access memories, mass storage devices, input interfaces, and output interfaces.

Claims

1. A method for training a neural network (RN) for detecting the limits of a traffic lane in which a vehicle is traveling at least partially autonomously, characterized in that: The training method comprises: - a first step (10-20), in which a first image of a first section of a traffic lane is acquired by a first sensor, and a second image of a second section of the traffic lane different from the first section is acquired by a second sensor, and each first image and each second image is recorded, an analysis vehicle simulating the passage of the at least partially automated vehicle on the traffic lane, wherein the first sensor and the second sensor are mounted on the analysis vehicle in different orientations, the first sensor is a so-called "frontal" sensor, and the second sensor is a so-called "body" sensor, - a second step (30-70), in which a so-called aerial image of the traffic lane is constructed in a top view by means of the recorded successive second images, the limits of the traffic lane are then determined from the aerial image, and elements capable of characterizing the limits of the traffic lane are determined from each first image, it is then determined whether these determined elements extend these determined limits, and in the affirmative case, each determined element is recorded in a database (BD), and in the negative case, each determined element is recorded in the database (BD) corresponding to the determined limits, and - A third step (80), in which the neural network (RN) is trained by means of records from the database (BD) so that it learns that, when there is an element just determined from the first image, it must either use the element when it is not stored corresponding to a limit, or use the limit stored corresponding to the element.

2. The training method according to claim 1, characterized in that: In the first step (10-20), a first image of a first successive segment of the traffic lane and a second image of a second successive segment of the traffic lane are acquired, wherein each of the first successive segments is located at a predefined distance in front of the analysis vehicle, and each of the second successive segments is located in the close vicinity of the analysis vehicle.

3. The training method according to claim 2, characterized in that: In the first step (10-20), a second image of a second, successive section of the traffic lane is acquired, the second, successive section being located in the immediate vicinity of the rear of the analysis vehicle.

4. The training method according to any one of claims 1 to 3, characterized in that: In the first step (10-20), each first image and each second image is recorded corresponding to at least one background information, and, in the second step (30-70), when the determined element extends the determined boundary, each determined element is recorded in the database (BD) corresponding to each recorded background information corresponding to each determined element, and in the negative case, each determined element is recorded in the database (BD) corresponding to the determined boundary and to each recorded background information corresponding to each determined element, and, in the third step (80), the neural network (RN) is trained by the records of the database (BD) so that the neural network learns that when there is at least one current background information and there is an element just determined from the first image, the neural network needs to either use the element when it is stored corresponding to the current background information, or use the boundary stored corresponding to the current background information and to the element.

5. The training method according to claim 4, characterized in that: In the first step (10-20), at least one background information selected from the following information is used: meteorological information related to the current environment of the analysis vehicle, light information characterizing the external brightness in front of the analysis vehicle, material occlusion information received in the analysis vehicle through the channel, activation of the external lighting system of the analysis vehicle, and the solar angular position of the sun relative to the horizon located in front of the analysis vehicle.

6. The training method according to claim 5, characterized in that: In the first step (10-20), the meteorological information is received into the analysis vehicle via a radio channel, or provided by a detector mounted on the analysis vehicle.

7. The training method according to claim 5 or 6, characterized in that: In the first step (10-20), the sun's angular position is determined from the acquired image, or is determined based on the current time and the current geographic location of the analysis vehicle.

8. A computer program product comprising a set of instructions capable of implementing, when executed by a processing unit, a training method according to any of the preceding claims for training a neural network (RN) for detecting limits of a traffic lane in which a vehicle is traveling at least partially autonomously.

9. A training device (DE) for training a neural network (RN) for detecting the limits of a traffic lane in which a vehicle is traveling at least partially autonomously, characterized in that The training device comprises at least one processor (PR) and at least one memory (MD), wherein the at least one processor and the at least one memory are configured to perform the following operations: - obtaining a first image of a first section of a traffic lane, the first image being acquired by a first sensor, and obtaining a second image of a second section of the traffic lane different from the first section, the second image being acquired by a second sensor, an analysis vehicle simulating the driving of the at least partially automated vehicle on the traffic lane, wherein the first sensor and the second sensor are mounted on the analysis vehicle in different orientations, the first sensor being a so-called "frontal" sensor and the second sensor being a so-called "body" sensor, and then - constructing a so-called aerial image of the traffic lane in a top view by means of the recorded successive second images, and then - determining the limits of the traffic lane from the aerial images and determining from each first image an element capable of characterizing the limits of the traffic lane, and then - determining whether the determined elements extend the determined limits and, in the affirmative, recording each determined element in a database (BD) or, in the negative, recording each determined element in said database (BD) in correspondence with the determined limits, and - The neural network (RN) is trained by means of records from the database (BD) so that it learns that, when an element is present which has just been determined from the first image, it must either use the element if it is not stored corresponding to a limit, or use the limit stored corresponding to the element.

10. A computer (OS) comprising a neural network (RN) to be trained for detecting the limits of a traffic lane in which an at least partially automated vehicle is traveling, characterized in that The computer further comprises a training device (DE) according to claim 9, which is connected to the neural network (RN).

Images