The invention relates to a headlight control system for a mo tor vehicle, comprising a controllable headlight adapted to generate variable illumination of the vehicle environment, an imaging apparatus adapted to capture images from a region in front of the motor vehicle, and a data processing device adapted to perform image processing of images captured by said imaging apparatus and to vary the light characteristics of said controllable headlight depending on said image pro cessing. The invention also relates to a method of training a machine learning model for such a headlight control system.

Systems for automatic headlight control typically base their decisions on explicit detection of individual vehicles or light sources. This makes it difficult to handle situations properly where vehicles appear in curves or over crests, or when headlights of oncoming vehicles are occluded by guard rails, barriers or other objects.

EP 2 495 127 B1 discloses a light aura detection approach to assist detection of vehicles appearing in curves or over crests. This approach, however, puts rather strict require ments on the appearance of such an aura and does not address the issue of other structures or objects occluding the light sources. Additionally, this approach must be complemented by a light source detector or vehicle detector. Furthermore, to control the headlights additional logic is needed for inter- preting the detections and to decide how the road ahead should be illuminated.

The problem underlying the present invention is to provide a holistic approach for a headlight control system adapted to handle different light conditions including all kinds of situ ations where structures and objects occlude light sources.

The invention solves this problem with the features of the in- dependent claims. According to the invention, a machine learn ing model is implemented in the data processing device which is trained to estimate and output an output signal represent ing a desired illumination of the vehicle environment from one or more images received as input from the imaging apparatus.

A human is able to interpret a multitude of light cues, such as halos, stray light or reflections, and thereby, for exam ple, determine likely locations of other vehicles. To mimic this approach, the invention provides a machine learning model which has been trained to holistically consider an entire im age, or a plurality of images, to directly estimate a head light control signal, or more generally an output representing a desired illumination of the vehicle environment. Generally, a desired illumination of the vehicle environment by one or more headlights includes the angular distribution and/or intensity distribution of all light sources present in the head light of a motor vehicle, including an illumination state of a high beam light source. In a preferred embodiment of the invention, the machine learn ing model is a convolutional neural network. The machine learning model is advantageously trained to output a desired illumination profile, which is preferably defined as any one of an upper vertical illumination angle per horizontal angular section, or a curve delimiting a desired area of illumination in an image, given an input image or several input images. In this case, the output signal output by the machine learning model advantageously comprises a desired illumination profile as defined above, expressed in angles relative to the optical axis of a headlight.

Ground truth data is preferably generated by manual annota- tion, i.e., by a human annotator. For distances significantly longer than the separation between imaging device (camera) and headlight, the desired illumination profile is closely approx imated by a curve delimiting the desired area of illumination in the image, which can easily be identified by a human anno- tator. Therefore, the machine learning model can advantageous ly be trained to output a curve delimiting a desired area of illumination in an image, rather than, for example, a desired illumination profile. Alternative output representations of the desired illumina tion, to be output by the machine learning model, may be uti lized. For example, the machine learning model may be trained to output a desired distance profile, which is preferably de fined as the illumination distance per horizontal angular sec- tion. In this case, the output signal output by the machine learning model advantageously comprises a desired distance profile as defined above.

In another embodiment, the machine learning model may be trained to output a desired per-pixel intensity map of the headlight. In other words, the output signal output by the ma chine learning model advantageously comprises a desired per- pixel intensity map. This output representation can be partic- ularly useful in combination with matrix headlights, wherein each headlight comprises a matrix of light sources, like LEDs, the intensity of which can be controlled individually. To make use of temporal information in the input, the machine learning model could take several sequential images as input, and/or exhibit internal feedback, e.g. as in a recurrent neural network. Preferably, the output signal is transmitted to a headlight controller generating a headlight adjustment signal, wherein preferably the headlight adjustment signal is fed back to the machine learning model. In other words, for improved perfor mance, the machine learning model may be provided with the current status of the headlights, which may be called an external feedback.

In a preferred embodiment, one might want to lower the inten sity or illumination angle of the headlights actively, to ena- ble detection of, for example, a halo from an oncoming vehicle that is occluded from view to the host vehicle by the terrain. The machine learning model could therefore be trained to iden tify situations where it is desirable to actively lower the intensity or illumination angle of the headlights, and to out- put an appropriate control or output signal for achieving this. The period of time for which the light intensity is lowered is preferably shorter than can be perceived by the driv er . If desirable, for example for reasons related to robustness or memory footprint, the machine learning model could preferably be adapted to process individually different parts of an input image in order to provide local contributions to the desired illumination, for example to the desired illumination profile, in said output signal from different parts of the field of view of the imaging apparatus . The invention also refers to a method of training a machine learning model for a headlight control system as described above. The machine learning model is trained in a supervised manner using a ground truth training data set specifying the desired output for each input sample, namely one image or a plurality of images, in the training data set.

Preferably, the ground truth training data set comprises one or more images, which are preferably captured by an existing imaging apparatus .

In an embodiment of the invention, ground truth data is generated by manual annotation of said one or more images. Prefera bly, the ground truth training data set comprises annotations in the form of a curve delimiting the desired area of illumi- nation.

Ground truth training data may be generated in several alter native or complementary ways. For example, the ground truth training data set may be generated using object detections from an existing object detection system, and/or an object tracker to track detected objects over consecutive images.

Since the object detections are used offline in this case, the objects may preferably be tracked in a non-causal manner to a point outside or beyond the initial detection range.

Alternatively, the ground truth training data set may be gen erated using semi-automatic annotations using output from an existing headlight control system or object detection system. In still another alternative embodiment, the ground truth training data set is generated using a recording of manual high beam control signals.

In the following the invention shall be illustrated on the ba sis of preferred embodiments with reference to the accompany ing drawings, wherein: Fig. 1 shows a schematic arrangement of a headlight control system;

Fig. 2 shows a flow diagram illustrating headlight control according the invention;

Fig . 3 shows an image with a curve delimiting a desired area of illumination; and

Fig . 4 shows a flow diagram illustrating training a machine learning model for a headlight control system.

The headlight control system 10 is to be mounted in or to a motor vehicle and comprises an imaging apparatus 11 for capturing images of a region surrounding the motor vehicle, for example a region in front of the motor vehicle. The imaging apparatus 11 may be mounted for example behind the vehicle windscreen or windshield, in a vehicle headlight, or in the radiator grille. Preferably the imaging apparatus 11 comprises one or more optical imaging devices 12, in particular cameras, preferably operating in the visible wavelength range, in the infrared wavelength range, or in both visible and infrared wavelength range, where infrared covers near IR with wave lengths below 5 microns and/or far IR with wavelengths beyond 5 microns. In some embodiments the imaging apparatus 11 comprises a plurality of imaging devices 12 in particular forming a stereo imaging apparatus 11. In other embodiments only one imaging device 12 forming a mono imaging apparatus 11 can be used.

The imaging apparatus 11 is coupled to an on-board data pro cessing device 14 adapted to process the image data received from the imaging apparatus 11. The data processing device 14 is preferably a digital device which is programmed or program mable and preferably comprises a microprocessor, a microcon troller, a digital signal processor (DSP) , and/or a micropro cessor part in a System-On-Chip (SoC) device, and preferably has access to, or comprises, a digital data memory 25. The data processing device 14 may comprise a dedicated hardware de vice, like a Field Programmable Gate Array (FPGA) , an Applica tion Specific Integrated Circuit (ASIC) , a Graphics Processing Unit (GPU) or an FPGA and/or ASIC and/or GPU part in a System- On-Chip (SoC) device, for performing certain functions, for example controlling the capture of images by the imaging appa ratus 11, receiving the electrical signal containing the image information from the imaging apparatus 11, rectifying or warp ing pairs of left/right images into alignment and/or creating disparity or depth images. The data processing device 14, or part of its functions, can be realized by a System-On-Chip (SoC) device comprising, for example, FPGA, DSP, ARM, GPU and/or microprocessor functionality. The data processing de vice 14 and the memory device 25 are preferably realised in an on-board electronic control unit (ECU) and may be connected to the imaging apparatus 11 via a separate cable or a vehicle da ta bus. In another embodiment the ECU and one or more of the imaging devices 12 can be integrated into a single unit, where a one box solution including the ECU and all imaging devices 12 can be preferred. All steps from imaging, image processing to possible activation or control of a safety device 18 are performed automatically and continuously during driving in re al time.

The invention is applicable to autonomous driving, where the ego vehicle is an autonomous vehicle adapted to drive partly or fully autonomously or automatically, and driving actions of the driver are partially and/or completely replaced or execut- ed by the ego vehicle.

The headlight control system 10 comprises one or more, for example two headlights 24 with at least one light source 20. Preferably, each headlight 24 is dynamically adjustable, i.e. the light profile of at least one light source 20 including the angular distribution and/or the intensity of the emitted light can be changed over time by an adjustment device 21 and controlled by a headlight controller 23. The headlight con troller 23 can be part of the processing device 14 or a sepa- rate processing device and part of the same ECU with the pro cessing device 14, or a different ECU. The imaging apparatus 11 is preferably directed in approximately the same direction as the headlights 24, such that the field of view of the imaging apparatus 11 and the illumination region of the headlights 24 at least partially overlap.

The adjustment device 21 may be adapted to adjust the corre sponding light source 20 in such a manner that the light beam or light cone 25 emitted by the headlight 24 is moved in a lateral direction and/or in a vertical direction or any other direction, as indicated by the arrows at the side of the light cones 25. The adjustment device 21 can be adapted to turn the complete headlight 24, to block or shield different parts of the light beam 25, to move one or more optical elements within the headlight 24, to change optical properties of one or more optical elements within the headlight 24, or any other suita ble mechanism. The adjustable headlight 24 may be an advanced lighting system, in particular based on LEDs, which can shape the light beam 25 around the oncoming vehicle without dazzling the oncoming driver.

The adjustment device 21 may be adapted to perform high beam control, i.e. to turn on and off the high beam comprised in the headlight 24 automatically as controlled by a high beam controller 26. The high beam controller 26 is preferably part of the headlight controller 23, but may also be a separate part .

As will be described in the following in more detail with re spect to Figure 2, the headlight controller 23 and/or the high beam controller 26, and thus the adjustment devices 21 for the headlights 24, are controlled during driving by the data pro- cessing device 14 on the basis of results obtained from data processing of the images received from the imaging apparatus

11. This is also called dynamic headlight control. Therefore, the data processing device 14 together with the headlight con troller 23 and/or the high beam controller 26 forms a dynamic headlight control device or controller.

According to the invention, a machine learning model 27, for example a convolutional neural network, is implemented in the processing device 14. The machine learning model 27 has been trained in a training phase prior to implementing it in the processing device 14, to directly estimate and output a headlight control signal and/or the desired illumination profile 29, from one or more entire images 28 received from the imag- ing apparatus 11 and input into the machine learning model 27. The training process will be described in more detail later with respect to Figure 4. During driving of the host vehicle, an image 28 or a plurality of images 28 captured by the imaging apparatus 11 are input to the machine learning model 27. The machine learning model 27 is capable of outputting an output signal 29 comprising a rep resentation of a desired illumination of the vehicle environ- ment by the headlights 24. The representation is for example a curve 41 delimiting a desired area of illumination in an image 40. As can be seen in Figure 3, the curve 41 is determined in a manner that the road 42, in particular the ego lane 43 and a region 45 at the road edge 44 of the ego lane 43, where pedes- trians, bicyclists and/or large animals may be expected, is well illuminated, however, without dazzling the driver of an oncoming vehicle 46. The curve 45 usually is an excellent ap proximation to the illumination profile, i.e. an upper verti cal illumination angle per horizontal angular section. Usual- ly, the region below the curve 41 can be well illuminated with high intensity, whereas the region above the curve 41 shall not be illuminated, or only with low intensity not dazzling the driver of an oncoming vehicle 46. The output signal 29 is forwarded to a headlight controller 23 which in turn sends a headlight control signal 30 to the head lights 24, in particular the headlight adjusting section 21 thereof. The headlight control signal 30 adjusts the head lights 24 in a manner that the region below the curve 41 is well illuminated and the region above the curve 41 is not, or only sparsely, illuminated. This includes possible automatic dim-out or switching off of the high beam in the headlights 12. The training of the machine learning model 27 is described in the following with respect to Figure 4. A large set of train ing images 50 is generated, for example with an imaging appa- ratus 51, which may preferably be similar or even identical to the imaging apparatus 11 of the vehicle where the machine learning model 27 is implemented. The set of training images 50 preferably covers an extensive range of light conditions which may occur in motor vehicle traffic.

All training images 50 are annotated by a human annotator 52. The annotations preferably comprise an appropriate illumina tion profile for every training image 50 as estimated by the human annotator 52. For example, the human annotator 52 may draw a curve 41 in every training image 50 such that the re gion below the curve 41 should be well illuminated and the re gion above the curve 41 is not, or only sparsely, illuminated, according to the best estimation of the human annotator. The set of training images 50 together with the corresponding an- notations, for example curves 41, forms an annotated or ground truth data set 53 of training images. The ground truth data set 53 of training images is input into the machine learning model 27 which can learn from this input to output an appro priate illumination profile for an arbitrary input image 28, or set of input images 28, when implemented in a motor vehicle .