Method for qualifying a trajectory hypothesis

Abstract

The invention relates to a method for qualifying a trajectory hypothesis (T), comprising the following steps: receiving a trajectory hypothesis (T), segmenting the trajectory hypothesis (T) into n portions (S), linearizing each of the n portions (S) into as many straight lines, extracting vanishing lines (F), determining a vanishing point (P), intersecting vanishing lines (F), calculating the orthogonal distance between the vanishing point (P) and the straight line linearizing the portion (S), and assigning a quality indicator to a portion (S), the quality being more favorable the smaller the orthogonal distance between the straight line linearizing the portion (S) and the vanishing point (P). Figure 1

Description

Title of the invention: Method for qualifying a trajectory hypothesis technical field

[0001] The invention relates, in the field of assisted or autonomous driving, to a method of qualifying a trajectory hypothesis. Previous technique

[0002] In an assisted or autonomous driving context, a trajectory hypothesis is a necessary entry point for multiple functions, such as route finding, time to collision calculation, etc.

[0003] The trajectory proposals or hypotheses originate from different estimating modules, based on different technologies and / or using different sensors. Depending on the origin and the sensors used, the confidence that can be placed in these estimating modules and their estimates varies considerably.

[0004] When the sensor data perception / fusion block does not contain a map information database, it can be difficult to choose the correct hypothesis from among all the proposals. This can, for example, be due to poor visibility caused by obstacles in the field of vision / perception: one or more other moving vehicles, parked vehicles, etc. This can temporarily, or depending on external conditions such as weather, lead to blindness or a decrease in the quality of an estimating module. This results in the transmission of numerous hypotheses that are of little or no relevance and must be processed despite their low level of confidence, whereas hypotheses of low relevance could advantageously be rejected as early as possible using a screening process based on a qualification of each hypothesis.

[0005] Also, the invention addresses this concern and proposes to qualify a trajectory hypothesis, based on an analysis of the environment as perceived by sensors, including, for example, a video camera. Description of the invention

[0006] The basic idea of ​​the invention is to compare a trajectory hypothesis with a vanishing point of the environment, as perceived by an image sensor, aiming preferentially towards the front of the vehicle.

[0007] The invention aims to provide a method, addressing this problem.

[0008] To this end, the invention relates to a method for qualifying a hypothesis of trajectory, including the following steps: - receipt of at least one trajectory hypothesis, - segmentation of said trajectory hypothesis into n portions, at least one hypothesis of trajectory. - linearization of each of the n portions into as many straight lines, - Environmental analysis to identify leakage paths, - determination of at least one vanishing point, intersection of vanishing lines, - for each pair of segment / vanishing point, calculate the orthogonal distance between the vanishing point and the line linearizing the segment, - for each pair portion / vanishing point, assignment of a quality indicator which is all the more favorable the smaller the orthogonal distance between the line associated with the portion and the vanishing point.

[0009] Specific features or embodiments, usable alone or in combination, are: - A trajectory hypothesis includes a curve, such as a clothoid or a polynomial curve, - the segmentation stage, cutting portions of actual length between 1 and 10 m, - The linearization step associates with the portion a tangent line at a point in the portion, a least squares line, or an interpolated line. - The extraction of vanishing lines is performed by image analysis, preferably from a video sensor, - only the upper part of the image is considered, preferably the upper two-thirds, - Determining a vanishing point requires a number of vanishing lines greater than a first threshold, and that these vanishing lines intersect within a spatial area of ​​radius less than a second threshold. - the orthogonal distance is measured between the vanishing point and its orthogonal projection onto the line associated with the portion, along said orthogonal projection.

[0010] According to another aspect, a data processing device implementing such a process. Brief description of the drawings

[0011] The invention will be better understood upon reading the following description, given solely by way of example, and with reference to the figures in the appendix in which:

[0012] [Fig-1] Figure [Fig.1] shows, in schematic view, an image from a video sensor, augmented with lines and vanishing point, compared to a hypothetical segmented trajectory projected onto said image,

[0013] [Fig.2] Fig.2 shows, schematically, the principle of calculating the distance orthogonal,

[0014] [Fig.3] Fig.3 shows, schematically, the principle of determining a vanishing point. Description of the implementation methods

[0015] The invention relates to a method for qualifying a trajectory hypothesis T. This method comprises the following steps.

[0016] In a first step, the method receives at least one trajectory hypothesis T in order to qualify it. One or more trajectory hypotheses T may be submitted. In the case of multiple trajectory hypotheses T, the method is repeated as many times as necessary for each trajectory hypothesis T. Each trajectory hypothesis T is provided in any possible form. It may be a set of successive passing points, possibly grouped or interpolated within a curve of any mathematical nature or form.

[0017] In a second step, each of said at least one trajectory hypothesis T is segmented into n portions S. A portion S is a piece of estimated trajectory T. The shape of a portion S depends on the shape of the estimated trajectory T. It may include several consecutive points, or a piece of curve.

[0018] In a third step, each of the n portions S is linearized by a straight line D associated with that portion S. This is repeated for all portions S of all trajectory hypotheses T, in order to produce as many straight lines D as there are portions S.

[0019] The fourth and fifth steps may possibly be carried out in parallel with the first three steps.

[0020] During the fourth step, more particularly illustrated in [Fig. 1], an analysis of the environment is carried out to extract leakage lines F.

[0021] During the fifth step, more particularly illustrated in [Fig.3], at least one vanishing point P is determined, an intersection of vanishing lines F among those previously extracted.

[0022] Then, subsequent to the third and fifth steps, during a sixth step, the orthogonal distance d between a portion S and a vanishing point P is calculated. This is carried out for each pair portion S / vanishing point P.

[0023] The principle of determining the orthogonal distance d, illustrated in [Fig.2], is to calculate the distance between the vanishing point P considered and the line D linearizing the portion S considered, according to an orthogonal projection.

[0024] In a seventh and final step, a quality indicator is assigned to each pair portion S / vanishing point P. This quality indicator is all the more favorable, indicating that portion S is more likely as part of a desirable trajectory hypothesis T, the smaller the orthogonal distance d is, as determined between the line D associated with portion S and vanishing point P.

[0025] It has been seen that a trajectory hypothesis T can be provided in any form: a set of points, a curve, etc. According to another characteristic, in the case where a trajectory hypothesis is a curve, this curve is preferably a clothoid or a polynomial curve. Such curve formats are classically used in the field, as they allow for a good representation of a continuous trajectory originating from the vehicle.

[0026] According to another feature, the segmentation step cuts portions S according to a criterion of actual length. Here, actual length is understood to mean the length, on the ground, as traveled by the vehicle. Advantageously, this actual length is between 1 and 10 m. This length can be reduced when the curvature increases and increased when the curvature decreases.

[0027] The linearization step can be performed using any method. According to one method, the linearization associates a line D tangent to the portion S at a point on the portion S. This point can be arbitrary: the midpoint, the first point, the last point, etc. According to another method, the linearization can also associate a line D with the portion S, passing as close as possible to the points of the portion S, previously decomposed into a set of points, the line D being the least-squares line with respect to this set of points. The line D associated with the portion S can also be a line determined by interpolation.

[0028] According to another feature, the extraction of vanishing lines F is performed by image analysis. This image preferably comes from a video sensor. The method for extracting vanishing lines F is assumed to be known to those skilled in the art. The principle consists of linearizing the image, for example using a Sobel filter, so as to make all the lines in the image appear. Among these lines, some are retained, for example based on considerations of minimum length and / or according to their orientation, a line that is not substantially radial having little probability of being a candidate for being a vanishing line F.

[0029] The vanishing line extraction step F searches for relevant lines, primarily lines derived from building alignments or street furniture. Therefore, the lower part of the image, being more likely to be cluttered with moving elements (other vehicles, pedestrians, etc.) or irrelevant elements in terms of vanishing lines, may prove less useful. Thus, in an optimization effort, according to another characteristic, it may be advantageous to perform the extraction only on the upper part of the image, which is potentially richer in relevant data. Only the upper part of the image is subjected to extraction. According to another characteristic, this upper part comprises the top two-thirds of the image.

[0030] It should be noted that the preceding criterion is environment-dependent. Thus, in an open urban environment, without parked vehicles, the sidewalk can prove to be a particularly advantageous source of vanishing lines. Similarly, in a motorway environment, the horizontal markings are a significant source of vanishing lines. In such environments, it may be preferable to consider the entire image.

[0031] According to another feature, more particularly illustrated in [Fig. 3], the determination of a vanishing point P is governed by selection criteria. Thus, according to a first criterion, a vanishing point P, in order to be retained as such, must be the point of intersection of at least a certain number of vanishing lines F, this number being at least equal to a first threshold. Furthermore, to be considered a point of intersection, a point must be located in a neighborhood, typically defined by a circle, whose radius is sufficiently small, i.e., less than a second threshold.

[0032] The first and second thresholds are preferably parameterizable, in order to be able to adapt the process.

[0033] As illustrated in [Fig. 3], a first bundle, here consisting of three vanishing lines Fl, is sufficiently convergent to intersect in a given neighborhood, represented by a circle surrounding the vanishing point PI, for the vanishing point PI to be considered and validated as a vanishing point PI. The same is true for the bundle of vanishing lines F2, which defines the vanishing point P2. The other intersections are not sufficiently consistent / close to define another vanishing point. The vanishing line F3, which is not angularly consistent with any other vanishing line F, is not retained.

[0034] As illustrated in [Fig. 2], the orthogonal distance d between the portion S, linearized by the line D, and the vanishing point P is measured as follows. The vanishing point P is projected orthogonally onto the line D, in an image P'. The orthogonal distance d between the portion S and the vanishing point P is equal to the distance separating the vanishing point P from its projection P'.

[0035] A trajectory hypothesis T is generally expressed in a vehicle frame of reference. Before calculating the distance d, it is necessary to place the trajectory hypothesis T / the portion S in the same frame of reference as the vanishing point P. For this, the portion S is advantageously projected onto the plane of the video image which enabled the detection of the vanishing points P, as illustrated in [Fig. 1].

[0036] The method according to the invention proposes a new approach to qualify a trajectory hypothesis T. It mainly uses buildings and street furniture to provide vanishing lines F. The quality of a trajectory hypothesis is a combination of the quality indicators of its portions S.

[0037] A combination of multiple leakage points P makes it possible to increase the relevance of a portion S, when this portion S obtains overall satisfactory quality indicators, when compared to several leakage points P.

[0038] Using different perception and calculation modules, several trajectory hypotheses T can be provided. The quality factors obtained by the segments S of each of these trajectory hypotheses T via the process according to the invention can allow the selection of the "best" trajectory hypothesis T, the one whose cumulative / average quality indicators of the S portions are the most favorable.

[0039] The invention can be applied in all fields using trajectory assumptions T: assisted or autonomous piloting, private vehicles, heavy goods vehicles, motorcycles, bicycles, etc.

[0040] The invention has been illustrated and described in detail in the drawings and the preceding description. This description is to be considered illustrative and given by way of example and not as limiting the invention to this single description. Numerous embodiments are possible.

Claims

Demands

1. A method for qualifying a trajectory hypothesis (T), characterized in that it comprises the following steps: - receiving at least one trajectory hypothesis (T), - segmenting said at least one trajectory hypothesis (T) into n portions (S), - linearizing each of the n portions (S) into as many lines (D), - analyzing the environment to extract vanishing lines (F), - determining at least one vanishing point (P), intersection of vanishing lines (F), - for each pair portion (S) / vanishing point (P), calculating the orthogonal distance (d) between the vanishing point (P) and the line (D) linearizing the portion (S), - for each pair portion (S) / vanishing point (P), assigning a quality indicator that is all the more favorable the smaller the orthogonal distance (d) between the line (D) associated with the portion (S) and the vanishing point (P).

2. A method according to claim 1, wherein a trajectory assumption (T) includes a curve, such as a clothoid or a polynomial curve.

3. A method according to any one of claims 1 or 2, wherein the segmentation step cuts portions (S) of actual length between 1 and 10 m.

4. A method according to any one of claims 1 to 3, wherein the linearization step associates with the portion (S) a tangent line at a point of the portion (S), a least squares line or an interpolated line.

5. A method according to any one of claims 1 to 4, wherein the extraction of leading lines (F) is carried out by image analysis, preferably from a video sensor.

6. A method according to claim 5, wherein only the upper part of the image is considered, preferably the upper 2 / 3.

7. A method according to any one of claims 1 to 6, wherein the determination of a vanishing point (P) requires a number of vanishing lines (F) greater than a first threshold, and that these vanishing lines (F) intersect in a spatial area of ​​radius less than a second threshold.

8. A method according to any one of claims 1 to 7, wherein the orthogonal distance (d) is measured between the vanishing point (P) and its orthogonal projection (P') on the line (D) associated with the portion (S), along said orthogonal projection.

9. Data processing device characterized in that it implements a method according to any one of the preceding claims.