Method for determining the three-dimensional geometry of a truck moving in the environment of a motor vehicle.

Abstract

Method for detecting trucks and modeling a three-dimensional geometry of the trucks detected in the environment of a motor vehicle. Method for detecting trucks and modeling the detected trucks in three dimensions, comprising iterations of the following steps: - acquiring at least one image from a camera (201) fitted to the motor vehicle (100), and detecting trucks in the at least one image including segmenting the at least one image to obtain segmented data, then - acquiring a point cloud from a lidar (202) fitted to the motor vehicle, and filtering the point cloud representing the trucks identified in the segmented data, the filtering allowing to obtain a point cloud instance for each truck identified, then - determining the contours of each truck identified, including determining the contour lines of each point cloud instance.Figure for the abridged version: 1.

Description

Title of the invention: Method for determining the three-dimensional geometry of a truck moving in the environment of a motor vehicle.

[0001] The invention relates to a method for determining the three-dimensional geometry of a truck moving within the environment of a motor vehicle. The invention further relates to a device for determining the three-dimensional geometry of a truck moving within the environment of a motor vehicle. The invention also relates to a computer program implementing the aforementioned method. Finally, the invention relates to a storage medium on which such a program is stored.

[0002] The detection of vehicles in surrounding traffic by an autonomous vehicle is an essential task. Autonomous vehicles analyze their environment to detect and categorize surrounding vehicles, particularly based on their geometric characteristics. However, this approach is problematic for types of vehicles whose shape and size vary considerably, as is the case, in particular, for trucks.

[0003] Approaches based on the use of three-dimensional neural networks are generally unsuccessful because traffic databases contain few truck instances, making it impossible for a neural network to generalize them correctly. Furthermore, these approaches are expensive to implement.

[0004] The object of the invention is to provide a method for identifying trucks and modeling the three-dimensional geometry of trucks detected in the environment of a motor vehicle that is simple to implement, in particular without resorting to complex identification techniques based on 3D neural networks, and reliable.

[0005] To this end, the invention relates to a method for detecting trucks and modeling a three-dimensional geometry of the trucks detected in the environment of a motor vehicle the motor vehicle being equipped with a set of sensors including a lidar and at least one camera, The process includes iterations of the following steps: - a step of acquiring at least one image from a camera, and detecting trucks in the at least one image, including the application of a segmentation algorithm to the at least one image in order to obtain segmented data, then - a step of acquiring a point cloud from the lidar, and filtering the point cloud representing the trucks identified in the segmented data, the filtering allowing to obtain a point cloud instance for each identified truck, then - a step of determining the contours of each identified truck, including determining the contour lines of each point cloud instance.

[0006] In one embodiment, the acquisition and detection step includes the implementation of an algorithm for tracking trucks detected in successive acquisitions.

[0007] In one embodiment, the dimensions of the determined contour lines are used to determine a length and / or width and direction of each identified truck.

[0008] In one embodiment, the truck detection method further includes a step of determining a most probable geometry of each detected truck, including a one-dimensional statistical analysis of the contour lines of the detected truck.

[0009] In one embodiment, the statistical analysis includes, for each given dimension taken from a detected truck length and / or a detected truck width and / or a detected truck height, - the construction of a histogram representing a distribution of measures of said given dimension in different classes, then - a determination of the most probable value of said dimension as being the value of the class of the histogram which groups the most measurements.

[0010] In one embodiment, the step of determining a most probable geometry of a truck includes the removal of detected contour lines when their length is less than a lower limit value, or greater than an upper limit value.

[0011] In one embodiment, the step of determining the most probable geometry of each detected truck includes: - the calculation of a confidence score / index associated with each measurement of the given dimension, and - a determination of the most probable value of the given dimension as being the measure associated with the highest confidence score.

[0012] In one embodiment, the confidence score associated with a given measurement of a given dimension of a truck is calculated as a function of: - a total number of measurements falling within the same histogram class as the given measurement, and / or - an estimated noise level relative to the point cloud representing the truck, and / or - a proportion of points located in the point cloud instance representing the truck and whose status is indeterminate, in particular whose status is neither free nor occupied.

[0013] The invention further relates to a device for detecting trucks and modeling the geometry of trucks detected in the environment of a motor vehicle equipped with a set of sensors including a lidar and at least one camera, the device comprising hardware and / or software elements implementing the method according to the invention.

[0014] The invention also relates to a computer program product comprising program code instructions recorded on a computer-readable medium to implement the steps of the process according to the invention when said program is running on a computer.

[0015] The invention also relates to a motor vehicle comprising a device according to the invention.

[0016] The attached drawings represent, by way of example, an embodiment of a truck detection device and modeling of the geometry of detected trucks according to the invention and an execution method of truck detection and modeling of the geometry of detected trucks according to the invention.

[0017] Fig. 1 schematically represents a motor vehicle equipped with a device for detecting trucks and modeling the geometry of detected trucks according to the invention.

[0018] Figure [Fig.2] defines a direct orthonormal frame of reference of the motor vehicle according to the invention.

[0019] Fig. 3 illustrates classes of objects that can be detected in the environment of the motor vehicle according to the invention.

[0020] Fig. 4 represents an image captured by a camera fitted to the motor vehicle according to the invention.

[0021] Fig. 5 represents an enclosing box of a truck located in the environment of the motor vehicle according to the invention.

[0022] Fig. 6 represents a plurality of bounding box detections.

[0023] Figure 7 represents a projection of points from the lidar into boxes encompassing areas determined from images from a camera.

[0024] Figure 8 illustrates a noise suppression treatment in a cluster of points from a lidar equipping the motor vehicle according to the invention.

[0025] Figure 9 illustrates a noise suppression treatment in a cluster of points from a lidar equipping the motor vehicle according to the invention.

[0026] Fig. 10 represents the axes of a lidar equipping the motor vehicle according to the invention.

[0027] Fig. 11 represents a top view projection of the axes of a lidar equipping the motor vehicle according to the invention.

[0028] Fig. 12 represents a top-view projection of lidar points determining the contours of a truck.

[0029] Fig. 13 represents two straight line segments defining the contour of a truck, the segments being obtained by applying the RANSAC algorithm to a set of points from the lidar.

[0030] Fig. 14 illustrates an anchor point and an end point of the segments defining a contour of a truck.

[0031] Fig. 15 illustrates an example of a histogram relating to the length of a given truck located in the environment of the motor vehicle according to the invention.

[0032] An example of a motor vehicle 100 equipped with a device 10 for detecting trucks and modeling the geometry of trucks detected in an environment of the motor vehicle 100 is described below with reference to [Fig.1].

[0033] With reference to [Fig.2], a first orthonormal frame R0 is defined directly (X0, Y0, Z0), the axis X0 being parallel to the longitudinal axis of the motor vehicle 100 and oriented towards the rear of the motor vehicle 100, the axis Y0 being transverse and directed towards the right of the vehicle, and the axis Z0 being vertical and oriented towards the top of the vehicle.

[0034] The motor vehicle 100 can be a vehicle of any type, for example a passenger vehicle or a utility vehicle.

[0035] The device 10 comprises a sensor array 20 including at least one camera 201 located at the front of the vehicle. Preferably, the sensor array comprises at least two cameras, or even at least three cameras. In one embodiment, the cameras 201 may be located at the front and / or rear of the motor vehicle 100. Each camera 201 generates two-dimensional RGB images of an area of ​​the environment located within the field of view of said camera.

[0036] The sensor assembly 20 further includes a lidar 202 which generates a three-dimensional N_3D point cloud. In a preferred embodiment, the lidar 202 sensor is a multiplane sensor covering at least sixteen planes.

[0037] Moreover, the range of the lidar 202 along a horizontal plane is greater than or equal to the range of the cameras 201 of the sensor set 20. In other words, the field of view of the cameras 201 is covered by the point cloud N_3D generated by the lidar 202.

[0038] Calibrations of the cameras 201 and the lidar 202 are carried out, in particular in order to be able to project a three-dimensional N_3D point cloud from the lidar 202 onto two-dimensional I_2D images from one of the cameras 201.

[0039] To this end, the cameras 201 are first individually calibrated; these are referred to as the intrinsic calibration parameters of the camera. In one embodiment, the intrinsic calibration is performed using an algorithm based on the pinhole model of the camera. The calibration of the intrinsic parameters of each camera 201 involves processing images, captured from different angles, of an object comprising a plurality of points with known coordinates, the object generally being a checkerboard pattern. This processing makes it possible to obtain, for each camera 201, a so-called camera matrix containing the focal length and optical centers, as well as distortion coefficients, which are then used to correct the previously obtained camera matrix in order to eliminate distortion errors.

[0040] Then the cameras 201 and the lidar 202 are calibrated together. An extrinsic calibration is performed between the cameras and the lidar so that the three-dimensional point cloud N_3D from the lidar can be projected onto a two-dimensional image I_2D from a camera. The extrinsic calibration parameters between the cameras 201 and the lidar 202 include the position and orientation of each sensor 201, 202, which allows a geometric relationship to be defined between the different sensors.

[0041] The device 10 further includes a microprocessor 4. The microprocessor 4 is capable of receiving and processing data from the cameras 201 and the lidar 202. In an alternative embodiment, the device 10 could include several microprocessors 4 and the data from the cameras 201 on the one hand, and from the lidar 202 on the other hand, can be processed on two separate microprocessors 4.

[0042] In one embodiment, the microprocessor 4 is connected to the sensors (cameras 201 and lidar 202) via a communication network, for example an Ethernet network.

[0043] Device 10, and particularly microprocessor 4, enables the implementation of a method for detecting and determining the geometry of a truck. It mainly comprises the following modules 41, 42, 43, 44: - a module 41 for acquiring at least one image from a camera, and for detecting trucks in at least one image, the module collaborating with at least one camera from the sensor array, - a module 42 for acquiring and filtering a point cloud from the lidar, the module collaborating with the lidar 202 of the sensor array, - a module 43 for contour search of each identified truck, then - a module 44 for determining the most probable geometry of each detected truck.

[0044] The motor vehicle 100, in particular the truck detection and truck geometry determination device 10, preferably comprises all the hardware and / or software elements configured to implement the process defined in the object of the invention or the process described below.

[0045] An embodiment of the method for determining the geometry of a truck is described below with reference to Figures 3 to 15. In the embodiment shown in Figures 3 to 15, the method comprises iterations over four steps E1 to E4.

[0046] In step El, at least one image is acquired from a camera, and trucks are detected in at least one image.

[0047] To this end, in step El, calculations are implemented to detect trucks represented in two dimensions in the image from camera 201. In particular, an instance segmentation based on the image is performed, that is to say, a processing of detection and categorization of the elements present in a scene represented by the image from the camera.

[0048] Instance segmentation differs from conventional object detection because instance segmentation predicts the pixel-level boundaries of each object, whereas object detection only predicts the approximate location of an object.

[0049] The so-called instance segmentation processing makes it possible to distinguish different instances of the same class and to treat them as different objects. Figure 3 illustrates three different classes of objects: a Class C1 concerning trucks, a Class C2 concerning passenger cars, and a Class C3 concerning pedestrians. In Figure 3, the truck class is represented by two instances, the passenger car class by one instance, and the pedestrian class by three instances. Trucks are notably represented by instance masks (also called bounding masks) that follow the detection contours and can be approximated by rectangular bounding boxes for ease of manipulation.

[0050] In the remainder of the document, a so-called bounding box of a truck is a contour of a flat surface representing the truck in an image from the camera.

[0051] In the remainder of this document, the term "instance" is used to refer to a computer representation modeling a given truck detected in the image from a camera. Two trucks detected as distinct in the image segmentation step are represented by two distinct instances.

[0052] In an advantageous embodiment, a tracking process is implemented for each instance, specifically for each truck instance. The tracking process aims to track the movement of each truck instance. An embodiment of the tracking process is described below.

[0053] During the segmentation processing of a first image, the boundaries (or bounding boxes) of each detected truck are determined and stored in memory. Then, an algorithm for tracking the movements of the detected trucks is implemented by processing a succession of images from camera 201. In particular, a detection cost matrix can be associated with each new image. Each cell of the cost matrix is ​​then defined as a calculated distance between the corresponding center of the bounding box of the current image and the center of the corresponding bounding box associated with an agent (during the previous tracking iteration). From the cost matrix thus obtained, each center of a bounding box of the current image is associated with the instance whose cost is the lowest, as long as this cost remains below a predefined threshold.Advantageously, instances that have not been associated with a bounding box during a given number of iterations are removed. Furthermore, new instances that were not detected during previous iterations are added.

[0054] The bounding box tracking process generates a detection table, each detection containing, for example: - the coordinates of a point located in the upper left corner of the bounding box of the detected truck, - the coordinates of a point located in the bottom right corner of the bounding box of the detected truck, - the coordinates of a center of the bounding box, - an agent identification number (also called "tracking ID") provided by the tracking process.

[0055] An example of determining a bounding box B4 is illustrated in Figures 4 and 5. Figure 4 is a two-dimensional image I_2D captured by camera 201, the image I_2D representing, in particular, a truck C4. Figure 5 represents the bounding box B4 associated with the truck C4 and the tracking identifier ID4 of the bounding box B4.

[0056] In step E2, point clouds from the lidar are acquired and filtered by projecting them into the bounding boxes previously defined in the first step EL. The filtering thus makes it possible to obtain a three-dimensional point cloud instance for each truck that was detected in step EL.

[0057] For example, [Fig. 6] represents a plurality of B6 bounding box detections defined in step 1E1 from one or more images from camera 201. [Fig. 7] represents a projection carried out in step E2 of an N_3D point cloud from the lidar onto the bounding boxes detected during step 1E1. This yields several point groups, some of which may be noise. For example, due to the lidar's range, objects located behind a truck, which constitutes noise. In other words, the noise comes from points located behind a truck, and the closest points always come from the truck.

[0058] In an advantageous embodiment, a clustering algorithm based on Euclidean distance is used to suppress noise, i.e. false detections.

[0059] In a two-dimensional coordinate system, the Euclidean distance between a point A with coordinates (xl, yl) and a point B with coordinates (x2, y2) is calculated according to the following formula: S = - Xi)2 4- (va — yi)2'

[0060] Since the clustering algorithm can produce several sub-clusters, a selection criterion is applied: only the sub-cluster closest to the recording system, i.e. the lidar 202, and comprising at least a minimum number of points is taken into account for the following steps.

[0061] Figures 8 and 9 illustrate the noise suppression process. In [Fig. 8], the measured distance DI between a cluster of points and a given first point is significant, specifically exceeding a given threshold. Therefore, the given first point will not be included in the cluster. Conversely, in [Fig. 9], the measured distance D2 between the cluster of points and a given second point is small, specifically below a given threshold. Therefore, the given second point will be included in the cluster.

[0062] As illustrated by Figures 10 and 11, a direct orthonormal coordinate system (XI, Yl, Zl) is associated with the lidar 202 and the second step E2 comprises: - a sub-step of measuring the height of the truck, by measuring the difference between the minimum and maximum values ​​of the lidar point cloud associated with truck detection; - a sub-step of projecting the three-dimensional points from the lidar 202 onto a plane containing the XI and Yl axes of the right-handed orthonormal coordinate system associated with the lidar. Indeed, since the height of the truck is known, it is no longer necessary to process the component along the Zl axis of the point cloud. The lidar point cloud is therefore transformed using a "bird's-eye view," that is, a projection onto a (XI, Yl) plane.

[0063] To this end, the point cloud is converted into a state called a "bird's-eye view," where the lidar points are viewed from above. To convert the points into this view, the Zl-axis coordinate is removed. The lidar points are thus represented in a two-dimensional image.

[0064] Fig. 12 thus illustrates a bird's-eye view N_3D_proj of the lidar points which were projected into the bounding box of a truck and then projected onto a plane (XI, Yl).

[0065] Then, from a bird's-eye view representation of the lidar points projected onto the bounding box, in step E3, contours of each identified truck are determined, each contour being defined in the (XI, YI) plane. Figure 13 illustrates an example of a truck contour constructed from the points shown in Figure 12. The truck contour is defined by two line segments S1 and S2, the first line segment S1 representing the length of the truck and the second line segment S2 representing the width of the box. The first and second segments together define a shape corresponding to the letter "L".

[0066] To this end, in step E3 it is possible to use the RANSAC algorithm (random SAmple Consensus), which is an iterative calculation method for estimating a mathematical model from a dataset that may contain outliers, in order to identify one or more straight line segments in the scatter plot. In the described embodiment, the mathematical model is previously defined as a straight line, defined by the equation of a straight line (y = mx + c).

[0067] The RANSAC algorithm selects a random subset of points and attempts to fit the model to the random subset of points. When the ratio between non-outliers and outliers exceeds a predefined threshold, the correct model has been found.

[0068] In implementing the invention, the RANSAC algorithm is applied first to define the first segment S1, and then a second time to define the second segment S2. The same algorithm can also be implemented to determine the truck's height. However, measurements along the Z-axis are less noisy than those along the X and Y axes, and the performance obtained by determining the truck's height directly in the filtered lidar point cloud is generally sufficient.

[0069] When the truck is only partially visible, the RANSAC algorithm can only determine one segment instead of two. In this case, it is not possible to obtain the complete dimensions: the length or width of the truck is missing. However, the missing dimensions can be recovered, for example by implementing an advantageous step of associating tracking identifiers with a histogram generated according to the method described below from successive acquisitions.

[0070] As illustrated by [Fig. 14], after determining the first segment SI and / or the second segment S2, the detected lines can be analyzed in order to to determine whether it is the length or the width of the truck, and to determine the truck's direction. One way to perform this analysis is to determine, for each segment: - a point PI closest to the lidar, called the anchor point, the anchor point being defined by its abscissa x_anchor and its ordinate y_anchor, - a point P2 furthest from the lidar, called the "extreme point", the extreme point being defined by its abscissa x_farthest and its ordinate y_farthest.

[0071] The lengths of the segments are obtained by measuring the distance between the anchor point and the extreme point.

[0072] From the coordinates of the anchor point PI and the extreme point P2, a direction of movement of the truck, also called "heading" or "cap", is calculated using the following formula: ■= tan x :------- ]

[0073] When a segment taken from the first or second segment SI, S2 is longer than a minimum length threshold (the minimum threshold being defined experimentally), said segment is classified as the longitudinal side of the truck; otherwise, it is classified as a width measured at the rear of the truck. The length and width of the truck are thus known.

[0074] Assigning identifiers to detections, and tracking detections on successive sensor acquisitions, makes it possible to refine measurements and find missing segments on certain acquisitions.

[0075] In systems comprising at least two cameras, the simultaneous detection of a truck by two cameras is likely to generate two truck detections with different tracking identifiers. To avoid this, each truck detection can trigger a check to ensure that the truck is not already identified by the other camera. When the system according to the invention comprises three cameras, including a front-facing camera (the front-facing camera being located at the front of the motor vehicle), then the truck identifier can, for example, be the one assigned to it by the front-facing camera.

[0076] In a 360-degree configuration using six cameras: - the front camera can assign the tracking identifiers of trucks detected by the cameras located at the front of the vehicle, including the front side cameras, and - the rear camera can assign the tracking identifiers of trucks detected by the cameras located at the rear of the vehicle, including the rear side cameras, with the front camera and the rear camera acting as so-called "reference" cameras.

[0077] Different identifier assignment criteria may be adopted for other configurations.

[0078] In an advantageous embodiment, all the points from a reference camera are coded, each point belonging to a truck being coded by a boolean whose value is "TRUE", the other points being coded by a boolean whose value is "FALSE".

[0079] The set of points thus coded from the field of view of the reference camera is compared to one or more sets of so-called "local" points, which are detected in the fields of view of one or more other cameras mounted on the vehicle. If a set of so-called local points and the set of points from the reference camera share at least a given number of common points, the local points and the reference points are considered to be associated with the same truck, and the tracking identifier that was assigned to the points from the reference camera is also associated with both detections of this truck.

[0080] In the advantageous step E4, the most probable geometry of each detected truck is determined.

[0081] To this end, step E4 may include the removal of detected contour lines whose dimensions are less than a lower bound or greater than an upper bound. These bounds may be defined by a legal framework, by knowledge of truck dimensions, or by experimentation. This filtering step makes it possible to remove unrealistic dimensions.

[0082] Advantageously, step E4 may include a statistical analysis of the length of the detected contour lines.

[0083] In particular, the statistical analysis may include, for each given dimension taken from a detected truck length and / or a detected truck width and / or a detected truck height, - the construction of a histogram representing a distribution of said given dimension, then - determining the most probable value of said dimension as being the most frequent value of the histogram.

[0084] An example of a histogram showing the length of a given truck located in the vicinity of motor vehicle 100 is illustrated in [Fig. 15]. The x-axis represents length values ​​expressed in meters, with the values ​​discretized in one-meter increments. The y-axis represents the number of occurrences of the different length intervals shown on the x-axis, obtained for the same truck over successive acquisitions. In the histogram shown in [Fig. 15], the most frequently measured length for the given truck is between 13 and 14 meters.

[0085] This statistical approach makes it possible to correct erroneous measurements of a given truck dimension caused by occlusions or other errors. Step E4 This allows the construction, for each detected truck and for each dimension of the detected truck, of a histogram that takes into account the measurement history of the length, width, or height of the truck identified by its tracking ID. The x-axis corresponds to a specific dimension expressed in meters (length, width, or height), and the y-axis represents the number of occurrences of a value for that specific dimension within a range of values. Because truck dimensions are expressed as real numbers, it is nearly impossible for the dimensions to be identical across two separate measurements, especially when the measurements are calculated from data from several different sensors. Therefore, the dimension values, expressed along the x-axis of the histogram, are discretized. Steps of 0.5 meters and one meter were tested.

[0086] This statistical approach also makes it possible to find the dimensions of a truck which could not be determined on a given acquisition, from all the acquisitions made at different times for the same truck.

[0087] Upstream or downstream of the construction of a histogram, two methods of calculating the dimensions of a truck can be used.

[0088] According to a first method, outliers of length and / or width and / or height are first eliminated by comparison to a minimum and a maximum threshold, these thresholds being, for example, thresholds defined by law. Unrealistic values ​​are thus eliminated. Next, the range of values ​​corresponding to the greatest number of occurrences is selected, and then the average of the values ​​in this range is used as the final value. This approach eliminates outliers caused by occlusions or detection errors. However, in certain circumstances, for example when a truck is always partially obscured, there is a risk that the most repeated value will come from an obscured detection, and the measured dimensions for that truck will then be incorrect.

[0089] In addition to the first method, or as an alternative to the first method, a second method can be applied, in which a note, corresponding to a confidence score, is associated with each given measurement of length, width or height.

[0090] A method for calculating a rating is described below. For clarity, the rating method applied to measuring the length of a given truck is described below.

[0091] In the described embodiment, the calculation of the score takes into account three parameters: - The initial score assigned to the given measurement of the length of the given truck is equal to the total number of measurements falling within the range of values ​​of the measurement given. For example, with reference to the histogram of [Fig. 15], the initial note 5 is assigned to a measured length of 8.5 meters. A noise level is estimated for each point cloud representing a truck. To estimate the noise of the point cloud associated with a given truck, a distance is calculated between each point in the cloud and the linear model found for the given truck in step E3 by applying the RANSAC algorithm. The number of points whose distance exceeds a given threshold and the number of points removed by applying the clustering algorithm in step E3 are used to define the noise level of the point cloud associated with the given truck. The higher the noise level, the more the score associated with the given measurement is penalized. The potential occlusion of a given truck detection is calculated by analyzing the "gray areas" of the point cloud. Gray areas indicate zones whose status (free or occupied) is undetermined, and for which no information from the lidar is available. This occurs, for example, when an object is located between the lidar and a given area, preventing lidar beams from reaching that area, which is then considered a gray area. The larger the gray area, the lower the score assigned to the given measurement.

[0092] This process thus makes it possible to associate a note with each given measure.

[0093] The measure obtaining the highest score is then selected.

[0094] At the end of step E4, an annotated video with a bounding box is obtained. three dimensions for each truck detected in the traffic surrounding the motor vehicle.

[0095] Finally, the process according to the invention has the following advantages.

[0096] First, it enables the detection of trucks in a point cloud from the lidar, and provides, for each truck detected: - the three-dimensional coordinates of the center of the truck's bounding box, - a heading of the truck, and - the dimensions of the encompassing box.

[0097] In addition, the method makes it possible to assign a tracking identifier to each detection to ensure temporal consistency of the generated images.

[0098] Furthermore, the system is capable of correcting the dimensions of the detected trucks based on information from all the processed images.

[0099] It should be noted that the processing implemented in the method according to the invention does not require the implementation of neural networks. The algorithms implemented do not need to be trained.

[0100] The method according to the invention makes it possible to increase the robustness of any process which relies on the detection of trucks.

[0101] The modular architecture of the process according to the invention makes it easy to improve any sub-task of the process, for example by using more complex statistical algorithms, to further refine the dimensions of the truck.

[0102] The method according to the invention can also be implemented "offline" on data from sensors (at least one camera and a Lidar) to determine the positions of truck bounding boxes, in order to automatically provide a ground truth usable for training neural networks configured to detect trucks and estimate their dimensions.

[0103] The method according to the invention has been defined to improve truck detection using an automatic labeling system. The tests carried out showed positive results; in particular, the method according to the invention proved to be more effective than equivalent systems based on neural networks.

Claims

Demands

1. A method for detecting trucks and modeling a three-dimensional geometry of the trucks detected in the environment of a motor vehicle (100) equipped with a sensor array (20) including a lidar (202) and at least one camera (201), the method being characterized in that it comprises iterations of the following steps: - a step (E1) of acquiring at least one image (I_2D) from a camera (201), and of detecting trucks (C1, C2, C3, C4) in the at least one image (I_2D) comprising applying a segmentation algorithm to the at least one image (I_2D) in order to obtain segmented data, then - a step (E2) of acquiring a point cloud (N_3D) from the lidar, and of filtering the point cloud representing the identified trucks (C1, C2, C3, C4) in the segmented data, the filtering enabling the obtaining a point cloud instance for each identified truck,then - a step (E3) of determining the contours of each identified truck, including determining the contour lines (SI, S2) of each point cloud instance.

2. A method for detecting trucks according to the preceding claim, characterized in that the acquisition and detection step (El) comprises the implementation of an algorithm for tracking detected trucks (Cl, C2, C3, C4) in successive acquisitions.

3. A method for detecting trucks according to any one of the preceding claims, wherein the dimensions of the determined contour lines are used to determine a length and / or width and direction of each identified truck.

4. A method for detecting trucks according to any one of the preceding claims, characterized in that it further comprises a step (E4) for determining a most probable geometry of each truck (Cl, C2, C3, C4) detected comprising a one-dimensional statistical analysis of the contour lines (SI, S2) of the detected truck.

5. A method for detecting trucks according to the preceding claim, characterized in that the statistical analysis comprises, for each given dimension taken from a length of a detected truck and / or a width of a detected truck and / or a height of a detected truck, - a construction of a histogram representing a distribution of measurements of said given dimension in different classes, then - a determination of the most probable value of said dimension as being the value of the class of the histogram which groups the most measurements.

6. A method for detecting trucks according to any one of claims 3 to 5, characterized in that the step of determining the most probable geometry of a truck includes the suppression of contour lines (SI, S2) detected when their length is less than a lower limit value or greater than an upper limit value.

7. A method for detecting trucks according to any one of claims 3 to 6, characterized in that the step of determining a most probable geometry of each detected truck comprises: - calculating a confidence score / index associated with each measurement of the given dimension, and - determining the most probable value of the given dimension as being the measurement associated with the highest confidence score.

8. Truck detection method according to the preceding claim, characterized in that the confidence score associated with a given measurement of a given dimension of a truck is calculated as a function of: - a total number of measurements located in the same class of the histogram as the given measurement, and / or - an estimated noise level relative to the point cloud representing the truck, and / or - a proportion of points located in the point cloud instance representing the truck and whose status is indeterminate, in particular whose status is neither free nor occupied.

9. Device for detecting trucks and modeling the geometry of trucks detected in the environment of a motor vehicle (100) equipped with a set of sensors (20) including a lidar (202) and at least one camera (201), the device comprising hardware and / or software elements implementing the method according to any one of claims 1 to 8.

10. Product computer program comprising program code instructions recorded on a computer-readable medium to carry out the steps of the process according to any one of claims 1 to 8 when said program is run on a computer.

11. Motor vehicle (100) comprising a device according to claim 9.