System and method for detecting objects in the immediate surroundings of a vehicle

Abstract

The invention relates to an object detection system (1) installed on board a vehicle, the system comprising a sensor (S1, S2, S3) and means (10) for detecting, using the sensor (S1, S2, S3), a plurality of objects, each detected object being associated with at least coordinates (xi, yi) as a function of time, the detection system (1) further comprising: - means (12) for calculating an uncertainty (Ui) for each object as a function of the coordinates (xi, yi) as a function of time associated with that object; - means (14) for selecting an object, from among the plurality of objects, as a function of the uncertainties (Ui) calculated for those objects; and - means (18) for modifying the position of the sensor (S1, S2, S3), which means are capable of adjusting the position of the selected object in a field of view of the sensor (S1, S2, S3).

Description

Description Title of the invention: System and method for detecting objects in the immediate environment of a vehicle

[0001] The present invention relates to the field of vehicle control systems, and more specifically concerns a method and system for detecting objects by sensors embedded in a vehicle.

[0002] Motor vehicles are increasingly equipped with advanced driver assistance systems (ADAS), or control systems when the vehicle is autonomous, using perception modules with one or more sensors, such as cameras, LiDAR (Light Detection and Ranging), or radar. Such a perception module analyzes the vehicle's environment sufficiently, using data provided by the sensors, to allow the vehicle equipped with this module to react to different driving situations, such as automatically performing certain maneuvers.

[0003] Vehicle sensors are generally mounted in locations that provide them with the widest and most relevant field of view possible, for example, by being placed on the front or rear of the vehicle. Despite these optimal locations, sensors may have blind spots that prevent them from detecting certain road users or situations important for driving, or they may be hindered by external interference such as reflections, water droplets, or dust on the windshield, for sensors located behind the windshield.

[0004] Furthermore, the vehicle's perception module analyzes objects located in the middle of the sensors' fields of vision more accurately than those located at the periphery of these fields of vision.

[0005] It is therefore understandable that the perception module used by advanced driver assistance systems or vehicle control systems does not always provide these systems with sufficient information, depending on the driving situation. Increasing the number of sensors in the vehicle compensates for this lack of information, but also increases the cost of the perception module.

[0006] The present invention aims to remedy at least in part the aforementioned drawbacks by providing a method and system for object detection that allows for a more precise analysis of the vehicle environment, in a low-cost manner.

[0007] To this end, the invention proposes a method for detecting objects, implemented in a vehicle equipped with at least one sensor, comprising the following steps: - detection, from data acquired by at least one sensor, of a plurality of objects, each detected object being associated with at least coordinates as a function of time, the detection method being characterized in that it further comprises the steps of: - Calculation of an uncertainty for each detected object, based on at least the time-dependent coordinates associated with that object. - selection of an object, from among a plurality of objects, based on the uncertainties calculated for these objects, and - modification of the position of at least one sensor, adjusting the position of the selected object within a field of view of said at least one sensor.

[0008] The detection of the plurality of objects is for example carried out by a perception module, receiving raw data from the sensor(s) and analyzing it by image processing or deep learning techniques, to identify the objects in the fields of vision of the sensors and to associate with each of these objects coordinates and possibly other elements such as a speed, an acceleration, a direction relative to the vehicle, a probability of existence, a width and a length.

[0009] Alternatively, each sensor includes both a hardware module for acquiring raw data (such as pixels), and an associated perception module.

[0010] The steps of the detection process according to the invention are of course preferably repeated, for example periodically, whenever advanced driver assistance systems or vehicle control systems are activated.

[0011] Thanks to the invention, the lack of information related to detected objects is quantified by an uncertainty calculated for each object. Then, the uncertainty of an object considered critical—for example, because it is not very visible and is close to the vehicle—is reduced by modifying the sensor's position relative to that object. Adjusting the object's position within the sensor's field of view is intended to reduce this uncertainty, for example, by fully positioning the object within the sensor's field of view or by centering the object within the sensor's field of view. The invention thus allows for a more detailed analysis of the vehicle's environment, relevant to its operation, and without significant additional cost.Indeed, simply rotating one of the vehicle's sensors is enough to center the selected object within the sensor's field of view, which is less expensive than using an additional sensor or repositioning the sensor relative to the vehicle. The invention can therefore be implemented simply, with minimal effort, by integrating software and a single actuator into the vehicle, which is less costly than expanding the vehicle's field of view by adding more sensors.

[0012] It should be noted that modifying the position of at least one sensor here refers, for example, to a rotation of the sensor on its own axis, or to a more significant movement of the sensor on the vehicle, such as a translation of the sensor along a rail fixed to the vehicle, depending on the implementation of the invention. Furthermore, when the detection step uses several sensors, the step of modifying the position of at least one sensor may consist of modifying the position of all these sensors or just one of them.

[0013] In one embodiment of the invention, the time-coordinates associated with each object correspond to parameters of a state vector linked to the object, the state vector being estimated by an extended Kalman filter.

[0014] The uncertainty associated with the object depends, for example, on the trace of the error covariance matrix of the extended Kalman filter. More simply, it depends, for example, on a sum of the variances of each parameter of the state vector.

[0015] The state vector characterizes the object by its position, but can also include a velocity, acceleration, orientation, dimensions, or probability of existence associated with the object. The sum of the variances of these parameters, which vary over time, allows us to quantify the uncertainty associated with the corresponding object. Of course, other types of calculations are possible.

[0016] Furthermore, in one embodiment of the invention, the step of calculating the uncertainty associated with each object takes into account a probability of existence of the object, assigned to the object during the detection step. This probability of existence is, for example, a value between zero and one, provided by a perception module linked to at least one sensor. The perception module uses, for example, several algorithms or several neural networks for this purpose, and compares their results to assign a probability of existence to each object detected by at least one of the two algorithms or neural networks. Methods for calculating such a probability of existence are known, for example, as "Deep Ensemble" and "Monte Carlo Dropout." The uncertainty of a detected object is, for example, reduced or even minimized in the calculation step when the probability of existence associated with that object is high.Indeed, a probability of existence close to the value of one indicates a good perception of the object by the sensors, and therefore, changing the position of the object in the field of vision of the sensor would not greatly improve this perception, perhaps to the detriment of other detected objects having a lower probability of existence.

[0017] On the other hand, in one embodiment of the invention, during the selection step, a weighted sum of the uncertainty associated with the object and the inverse of a distance from the object to Fau minus a sensor is calculated for each object, the selected object being the one with the largest weighted sum. This embodiment of the invention This allows the selection of the object whose detection will then be improved, based on a relevance criterion linked not only to the object's lack of visibility to the sensors, but also to the object's distance from the vehicle. It is understood that the closer and less visible the object, the more urgent it is to analyze it more thoroughly.

[0018] In this embodiment of the invention, the weighted sum may take into account a classification of the object, associated with the object during the detection step. For example, the weighted sum is increased when the detected object is classified as a human being or an animal, so that it is given priority for better detection by the sensors during the step of modifying the position of at least one sensor. This takes into account the vulnerability of the detected objects.

[0019] Furthermore, since the duration of the step to modify the position of at least one sensor is estimated beforehand, this step includes, for example, a substep of predicting the position of the selected object and the position of the vehicle at the end of the estimated duration, and a substep of reorienting the sensor so that the predicted position of the selected object within the sensor's field of view at the end of the estimated duration reduces the uncertainty associated with the selected object. Because sensor reorientation takes some time, during which the vehicle and the selected object may change position relative to each other, it is indeed preferable to predict the position of the selected object relative to the vehicle before reorienting the sensor, so that this reorientation correctly adjusts the selected object within the sensor's field of view.

[0020] During the reorientation substep, the sensor's viewing angle is, for example, modified so that the predicted position of the selected object corresponds to the center of the sensor's field of view at the end of the estimated time. Centering the selected object within the sensor's field of view generally reduces the uncertainty associated with that object. If this is not the case, the method can, for example, include another sensor position modification step, in which, during the reorientation substep, the sensor's viewing angle is modified so that the predicted position of the selected object is slightly offset from the center of the sensor's field of view, for example, by 20°. Several successive offsets can thus be tested until the uncertainty associated with the selected object is reduced.In order to determine whether the uncertainty associated with the selected object decreases, the detection process includes, for example, a step of calculating the uncertainty associated with the selected object after a new detection step and a new step of calculating the uncertainties associated with the detected objects.

[0021] Furthermore, the step of modifying the position of at least one sensor is, for example, followed by a step of resetting the position of at least one sensor. The reset position of the sensor corresponds to its default position when the invention is not in use. This minimizes or eliminates any change in the operation of advanced driver assistance systems or vehicle control systems that utilize the vehicle's sensors, compared to their operation without the detection method according to the invention. While it may be necessary to update these systems with the modified sensor position, preferably the sensor position remains modified from its initial position for only a very short time, depending in particular on the vehicle's speed. For example, when the uncertainty associated with the object is sufficiently small, the change in the sensor position is on the order of a few milliseconds.

[0022] The invention also relates to an object detection system, embedded in a vehicle, the detection system comprising at least one sensor and means for detecting, from data acquired by at least one sensor, a plurality of objects, the detection means being capable of associating with each detected object at least coordinates as a function of time, the detection system being characterized in that it further comprises:

[0023] - means of calculating an uncertainty for each detected object, based on at least the time-dependent coordinates associated with that object, - means of selecting an object, from among a plurality of objects, based on the uncertainties calculated for these objects by the computing means, and - means for modifying the position of at least one sensor, capable of adjusting the position of the selected object within a field of vision of said at least one sensor.

[0024] The object detection system according to the invention has advantages similar to those of the object detection method according to the invention.

[0025] Other features and advantages of the invention will become apparent from the following description on the one hand, and from several illustrative and non-limiting examples of embodiments given with reference to the attached schematic drawings on the other hand, in which:

[0026] [Fig. 1] represents an object detection system according to the invention, in one embodiment of the invention,

[0027] [Fig.2] schematically represents a vehicle equipped with the object detection system of [Fig.1], and

[0028] [Fig.3] represents an object detection method according to the invention, implemented by the object detection system of [Fig.1], in one embodiment of the invention.

[0029] As shown in [Fig. 1], an object detection system 1 according to the invention is mounted in a vehicle. It comprises at least three sensors, such as, for example: - a first SI sensor, which is a camera arranged at the top of the vehicle's windshield, - a second sensor S2, which is a radar sensor, arranged for example in the front bumper of the vehicle, and - a third sensor S3, which is a LiDAR sensor, arranged for example on a part of the roof of the vehicle, proximal to the front of it.

[0030] In this embodiment of the invention, actuators 18 are associated with each of the sensors SI to S3 to modify their orientation and therefore their field of view. These actuators 18 are, for example, electric motors capable of rotating each of the sensors SI to S3.

[0031] Of course, the vehicle can have fewer or many more sensors depending on its level of autonomy. This embodiment only mentions sensors located at the front of the vehicle, but the vehicle can certainly have sensors at the rear and sides, such as radar sensors, or additional cameras and LiDAR. One of the sensors could even be a drone. Only some of these sensors may be equipped with actuators capable of changing their positions, in order to limit the cost of the detection system.

[0032] The sensors SI to S3 are connected via a computer bus (for example, a CAN bus, short for Controller Area Network) to detection means 10, which consist of a perception software module, commonly used in autonomous vehicles, and for example, stored in a dedicated computer within the vehicle. This perception software module uses data from all the sensors SI to S3 to detect objects i in the vehicle's immediate environment and model a representation of this immediate environment. In this application, the immediate environment refers to an area around the vehicle with a radius of approximately 200 meters. The objects i include the road, signage, road markings, and obstacles on the road, such as other vehicles, pedestrians, etc. In this representation, each detected object i is associated with a specific piece of information.The objects i that move in the vehicle's environment are each associated with: - a relative position of object i with respect to the vehicle carrying the detection system 1, this relative position corresponding to a central point of object i, with coordinates x i5 y ; in an orthonormal coordinate system (Ot, Xt, Yt) shown in [Fig. 2]; this orthonormal coordinate system (Ot, Xt, Yt) is fixed relative to the vehicle and follows the vehicle's trajectory, the x-axis Xt of this system being tangent to this trajectory, and the y-axis Yt of this system being orthogonal to the x-axis Xt while remaining, like the x-axis Xt, parallel to the road on which the vehicle is traveling; the x-coordinate ; is an abscissa of the object i in this coordinate system, and the y coordinate ; is a ordinate of object i in this frame of reference; of course the position of object i can be expressed by other types of coordinates, or in a frame of reference not linked to the vehicle; - an absolute velocity of object i expressed by vx values ; and vy ; in the orthonormal coordinate system mentioned above; the value vx ; is the projection of this absolute velocity onto the x-axis of the orthonormal coordinate system and the value vy ; is the projection of this absolute speed onto the ordinate axis Yt of the orthonormal frame; alternatively, the perception module delivers a relative speed with respect to the vehicle; - an absolute acceleration of object i; alternatively the perception module delivers a relative acceleration with respect to the vehicle; - a length L ; and a width Wi as perceived in the orthonormal coordinate system (Ot, Xt, Yt); - a classification of object i, indicating a type of object, for example a vehicle, a traffic sign, a pedestrian, etc.; - a relative orientation 0 ; corresponding to the angle made by a straight line connecting the center of the vehicle to the object i, with respect to the x-axis Xt of the orthonormal coordinate system (Ot, Xt, Yt); and - a probability of existence P ; of object i.

[0033] It should be noted that for visibility reasons, some parameter letters are indexed in the text but not necessarily in the figures.

[0034] Since the perception module operates continuously when advanced driver assistance systems or vehicle control systems are activated, the values ​​of the various parameters mentioned above are continuously output from the detection means 10, i.e., in calculation steps. Most of these parameter values ​​vary over time, and therefore at each calculation step, whether due to the movements of the vehicle or the corresponding detected object, or to the imprecision of the detection algorithms of the detection means 10.

[0035] The perception module uses known techniques such as fusion of data from the vehicle's SI to S3 sensors, deep learning, classification algorithms, etc.

[0036] The probability of existence P ;of an object i is a value between 0 and 1, provided by the perception module on the basis of an uncertainty estimate provided for example by a neural network trained to estimate the probability of existence of each object detected by the perception module.

[0037] Other methods such as "Deep Ensembles" or "Monte Carlo Dropout" can be used to obtain this probability of existence. In particular, when the detection module uses multiple neural networks, each neural network being Used to detect objects in the vehicle's immediate surroundings, the Deep Ensembles method allows for the comparison of results provided by these neural networks to estimate the probability of an object's existence. Thus, if one neural network detects a pedestrian but the pedestrian is not fully visible to the sensors, another neural network might detect a cyclist instead. In this case, the Deep Ensembles method will assign a low probability of existence to the object detected as either a pedestrian or a cyclist, using, for example, a measure of mutual information. Different positions estimated by two separate neural networks for the same object at the same time will also result in a low probability of existence, based on the distance between these estimated positions.

[0038] It should be noted that the probability of existence P ;is an uncertainty estimate provided by the perception module of the detection means 10, distinct from an uncertainty calculation such as that performed by calculation means 12 for an uncertainty for each detected object i, these calculation means 12 being part of the detection system 10 according to the invention. This uncertainty calculation is in fact specific to the invention.

[0039] The computing means 12 receive the values ​​of at least some of the parameters, as provided as outputs from the detection means 10, to calculate, for each object i, an uncertainty U; associated with that object i. This uncertainty U; quantifies a lack of information related to that object i, requiring action on the sensors SI to S3. The details of this calculation will be described later in relation to a detection method 100 implemented by the detection system 1.

[0040] The detection system 1 also includes means 14 for selecting an object j from among all the objects i detected by the detection means 10, based in particular on the uncertainties U ; objects i but also classifications of objects i. Alternatively, the means of selecting an object i 14 do not use the classifications Ci of objects i. These means of selecting 14 make it possible to prioritize the objects for which to fill the lack of information acquired for these objects by the detection means 10, in particular according to the urgency of filling this lack of information, as will be described later in relation to the detection process 100 implemented by the detection system 1.

[0041] Finally, the detection system 1 includes a module 16 for determining at least one angle kThe orientation of one of the sensors SI at S3 is modified, this angle being determined so as to adjust the position of the object j selected by the selection means 14, within the field of view of the sensor in question. This adjustment allows for the acquisition of more information about the object j and thus reduces the uncertainty Uj as subsequently recalculated by the calculation means 12. This adjustment therefore consists, at a minimum, of modifying the position of the object j within the field of view of a sensor, while maintaining This object j is placed within the sensor's field of view or moved further into it. Specifically, the centering of object j within the sensor's field of view can be a default setting for the object j's position within the sensor's field of view.

[0042] It is understood that the computing means 12 operate continuously, like the detection means 10, in calculation steps. The same applies to the selection means 14 and the determination module 16. The computing means 12, the selection means 14, and the determination module 16 are, for example, implemented in software within a vehicle computer, which may be separate from the computer housing the perception module.

[0043] Module 16 of determination having determined one or more angles k then commands a rotation, according to this angle or these angles, of one or more of the actuators 18, capable of rotating one or more of the sensors SI to S3. For example, if the sensor S1 is involved in an angle k determined by the determination module 16, the corresponding actuator rotates the sensor SI by this angle karound an axis of rotation of the sensor SL If the sensor SI is fixed to the vehicle by means of a ball joint, the determination module 16 can determine several angles of rotation, corresponding to as many axes of rotation, to modify the position of the sensor SL The determination module 16 can also determine a distinct angle for each of the sensors SI to S3 or for two of them, so that respectively each of the sensors SI to S3 or two of them then have their positions modified by all or part of the actuators 18.

[0044] We now describe in relation to [Fig.3], the detection process 100 as implemented by the detection system 1.

[0045] A first step 102 of the detection process 100 is a detection step, implemented by the detection means 10, of a plurality of objects i in the immediate environment of the vehicle, based on data acquired by sensors S1 to S3. In this first step 102, the detection means 10 provide, as output, for each detected object i, the values ​​of the parameters described above, namely, in particular, the x coordinates. ; , y ; of object i and its relative velocity expressed in the orthonormal coordinate system (Ot, Xt, Yt), the length L ; and the width W ; as perceived by the detection means in this orthonormal coordinate system, the C classification ; of object i, its relative orientation 0 ; , and its probability of existence P ; .

[0046] A second step 104 of the detection process 100, following the first step 102, is the calculation of an uncertainty U ;for each object i detected during the first step 102. This calculation is performed by the computing means 12 of the detection system 1. This second step 104 includes the application of an extended Kalman filter to the state vector:

[0048] The index t expresses the time dependence of this state vector. Using this extended Kalman filter allows us to estimate the state of object i based on the values ​​provided by the detection methods at time t, as well as those from previous times, and this in real time. The outputs of the detection methods are treated as measurements. In a known manner, such a filter analyzes the history of measurements associated with a state vector to recursively predict the state vector at the considered calculation step, for example, corresponding to time t. From this prediction and the measurement corresponding to this calculation step, the filter estimates the values ​​of the state vector parameters at time t. This estimation provides a value that is more accurate than the measurements at time t because it eliminates measurement noise, related, for example, to vehicle dynamics and sensor latency.

[0049] Therefore, after applying the extended Kalman filter in this second step 104, we obtain an estimated state vector x ct a covariance matrix P t representing the error in the accuracy of the state vector X estimation t The covariance matrix P t is also called the error covariance matrix in Kalman filter theory.

[0050] It should be noted that the extended Kalman filter is non-linear due to the evolution model of the state vector X t this one is based for example on a bicycle model of the vehicle, in which the steering angle ô, represented [Fig.2], is considered constant and the speed V of the vehicle, represented [Fig.2], is also considered constant.

[0051] The estimated state vector x is sent over the vehicle's computer bus and is used in particular by the vehicle's advanced driver assistance systems.

[0052] The trace of the covariance matrix P t , which is equal to the sum of the variances of the parameter values ​​of the state vector, provides a measure of the accuracy of the detection means 10 and thus quantifies an uncertainty Ui for object i, in the case where the probability of existence P ; of this object i is below a predetermined threshold.

[0053] Indeed, in this embodiment of the invention, an uncertainty Lf is defined for each detected object i, as a function of the variances of its state parameters and the probability of existence P ; of this object i. It is understood that other ways of defining this uncertainty are conceivable for each object i, for example as a function of more state parameters than those of the state vector X t , or without taking into account the probability of existence P ;of object i. Furthermore, in an alternative embodiment of the invention, the uncertainty calculation step 104 does not use an extended Kalman filter, but instead calculates only the sum of the variances of the parameters of the state vector X t .

[0054] More specifically, in this embodiment of the invention, the uncertainty U; is calculated as follows:

[0055] U; = tr(P t )*b with b=0 if P ; >th or b=l otherwise,

[0056] Or :

[0057] -tr (P t ) is the trace of the covariance matrix P t , - b is a boolean, and - th is the predetermined probability threshold of existence above which the detection of object i is considered good enough not to require an adjustment of the position of sensors SI to S3. This predetermined threshold is a value strictly between zero and one.

[0058] For example, if the predetermined threshold th is equal to 0.5: - the uncertainty U ; of an object i with probability of existence P ; equal to 0.3 will be equal to the trace of the covariance matrix P t calculated for this object i, and - the uncertainty U of an object i with probability of existence P ; equal to 0.95 will be equal to zero.

[0059] The predetermined threshold th prevents the consideration of an object i in the subsequent step of selecting an object i for which uncertainty needs to be reduced, when the detection of that object i is already sufficiently accurate. This threshold therefore avoids modifying the position of sensors SI to S3 to target objects, resulting in a very limited improvement in understanding the vehicle's immediate environment.

[0060] A third step 106 of the detection process 100 is then the selection of an object j, from among the plurality of objects i detected, according to the uncertainties U ; calculated for these objects i. This selection is implemented by the selection means 14 of the detection system 1.

[0061] More specifically, in this third step 106, the following weighted sum is calculated for each detected object i: - di is the Euclidean distance from object i to the vehicle, calculated as a function of the x coordinates ; and y ; of object i, -w3 is a positive coefficient based on the classification Ci of object i, larger the more vulnerable the object i is, for example a pedestrian or a cyclist; -wl is a positive constant weighting the importance given to the uncertainty U of object i relative to its proximity to the vehicle; and

[0064] -w2 is a positive constant weighting the importance of the proximity of object i to the vehicle, relative to the uncertainty of object i.

[0065] The object j selected in this third step 106 is then the one with the largest weighted sum calculated in this third step 106. This selection allows us to determine the object whose detection uncertainty most urgently needs to be reduced. It is therefore a function of the uncertainty U ; previously calculated, but also the proximity of object i to the vehicle and the vulnerability of object i. Of course, in variants, more or fewer criteria are taken into account, the main criterion being the uncertainty U; calculated during the second step 104.

[0066] A fourth step 108 of the detection method 100 is then the modification of the position of at least one of the sensors SI at S3 so as to decrease the uncertainty Uj of the object j, as recalculated by the computing means 12 following this modification of the position of the sensor(s) SI at S3. For simplicity, it is assumed in this embodiment of the invention that only the position of sensor SI is modified. Furthermore, since this fourth step 108 is of non-negligible duration compared to the vehicle's dynamics, it takes into account an estimated time At required for the reorientation of sensor SI. This estimated time At is, for example, predefined for each type of sensor and is set here, for example, to a few milliseconds.

[0067] The fourth step 108 then comprises a first substep of prediction 1082 of the position Xj,yj of the selected object j and of the position of the vehicle, at the end of the estimated duration At, necessary for the modification of the position of the sensor SI, which here is a rotation of the sensor SI on itself, around a vertical axis, by an angle k to be determined, this angle k being represented in [Fig.2]. The vertical axis is orthogonal to the axes of the abscissa and ordinate of the orthonormal coordinate system (Ot, Xt, Yt) fixed to the vehicle.

[0068] This first prediction substep 1082 uses the equations of the dynamics of the vehicle and the selected object j to determine the position of object j relative to the vehicle at the end of the estimated time At.

[0069] Then a second sub-step 1083 of this fourth step 108 of modifying the position of the SI sensor, is the calculation of the angle kof rotation of an optical axis x of the sensor SI, shown in [Fig. 2], with respect to the x-axis Xt of the orthonormal coordinate system (Ot, Xt, Yt) fixed to the vehicle, this angle Q k corresponding to an ideal angle at which the SI sensor should be positioned to minimize the uncertainty Uj of the selected object j. This calculation takes into account the current position of the SI sensor's optical axis x, and the previously predicted position of object j relative to the vehicle. Of course, this angle k is calculated so as to remain within the calibration limits of the SI sensor.

[0070] When the selected object j is not centered in the field of view of the SI sensor, the angle k is, for example, calculated so that the optical axis x of the SI sensor passes through the predicted position of object j at the end of the estimated time At. When the selected object j is already centered in the field of view of the SI sensor, the angle kis for example calculated so that the optical axis x of the sensor SI is offset by a predefined angle from the predicted position of the object j at the end of the estimated time At.

[0071] The first prediction substep 1082 and the second calculation substep 1083 of the angle k are performed by the determination module 16 of the detection system 1.

[0072] Finally, a last sub-step 1084 of this fourth step 108 of modifying the position of the SI sensor, is the rotation of the SI sensor by the angle k calculated previously, by at least one of the actuators 18, provided that this angle k is greater than a minimum angle, for example set at 1 degree. This reorientation of the SI sensor therefore does not take place for a minimum reorientation of the SI sensor which would result in a minimal gain in the accuracy of the detection of the selected object j.

[0073] At the end of this last substep 1084, the new orientation of the sensor SI, i.e. the new position of its optical axis x, is provided to the detection means 10 so that they can update the values ​​of the parameters and uncertainties related to the detected objects in the next calculation steps.

[0074] The four steps 102 to 108 of the detection process 100, as described above, are repeated continuously one after the other. However, the detection process 100 may include a step 109 following each fourth step 108 of the detection process 100, in which a number of iterations n of the fourth step 108 is compared with a predefined number of reorientations of at least one of the sensors SI to S3, for example, equal to two.

[0075] In the case where the number of iterations n of the fourth step 108 is strictly greater than this predefined number (branch Y), the detection process then includes, for example, a step 110 of resetting the position of the sensors SI to S3, in order not to disrupt the normal operation of the advanced driver assistance systems which use the outputs of the detection means 10. Otherwise, in the case where the number of iterations n of the fourth step 108 is less than or equal to the predefined number (branch N), the fourth step 108 loops back to the first step 102 of the detection process 100.

[0076] Of course, the invention is not limited to the examples just described, and many modifications can be made to these examples without departing from the scope of the invention. In particular, the characteristics of the different embodiments can be modified. of the invention envisaged in this application, may be combined to realize the invention, insofar as these variants are not incompatible with each other.

Claims

Demands

1. A method for detecting objects (100), implemented in a vehicle equipped with at least one sensor (S1, S2, S3), comprising the steps of: - detection (102), from data acquired by at least one sensor (S1, S2, S3), of a plurality of objects, each detected object being associated with at least coordinates (x i5 y ; ) as a function of time, the detection method (100) being characterized in that it further comprises the steps of: - calculation (104) of an uncertainty (U ; ) for each detected object, based on at least the coordinates (x ; , y ; ) depending on the time associated with this object, - selection (106) of an object, from among the plurality of objects, according to the uncertainties (U ; ) calculated for these objects, and - modification (108) of the position of at least one sensor (SI, S2, S3), adjusting the position of the selected object in a field of view of said at least one sensor (SI, S2, S3).

2. A method for detecting (100) objects according to claim 1, wherein the coordinates (x ; , y ; ) correspond to parameters of a state vector linked to the object, the state vector being estimated by an extended Kalman filtering.

3. A method for detecting (100) objects according to claim 2, wherein the uncertainty (U ; ) related to the object depends on the trace of the error covariance matrix of the extended Kalman filter.

4. A method for detecting (100) objects according to any one of claims 1 to 3, wherein the step of calculating (104) the uncertainty (U ; ) linked to each object takes into account a probability of existence (P ;) of the object, associated with the object during the detection step (102).

5. A method for detecting (100) objects according to any one of claims 1 to 4, wherein, during the selection step (106), a weighted sum of the uncertainty (U) is calculated for each object. ; ) related to the object and the inverse of a distance from the object to at least one sensor (SI, S2, S3), the selected object being the one with the largest weighted sum.

6. A method for detecting (100) objects according to claim 5, wherein the weighted sum takes into account a classification (Ci) of the object, associated with the object during the detection step (102).

7. A method for detecting (100) objects according to any one of claims 1 to 6, wherein a duration of the step of modifying the position (108) of at least one sensor (SI, S2, S3) being previously estimated, the step of modifying the position (108) of at least one sensor (SI, S2, S3) comprises a substep of predicting (1082) a position of the selected object and a position of the vehicle at the end of the estimated duration, and a substep of reorienting (1084) the at least one sensor (SI, S2, S3) such that the predicted position of the selected object in the field of view of the sensor (SI, S2, S3) at the end of the estimated duration decreases the uncertainty (Ui) related to the selected object.

8. A method for detecting (100) objects according to claim 7, wherein during the reorientation substep (1084), the viewing angle of the sensor (SI, S2, S3) is modified so that the predicted position of the selected object corresponds to the center of the field of view of the sensor (SI, S2, S3) at the end of the estimated time.

9. Method for detecting (100) objects according to any one of claims 1 to 8, wherein the step of modifying (108) the position of at least one sensor (SI, S2, S3) is followed by a step of resetting (110) the position of at least one sensor (SI, S2, S3).

10. Object detection system (1), mounted in a vehicle, the detection system (1) comprising at least one sensor (SI, S2, S3) and detection means (10), from data acquired by the at least one sensor (SI, S2, S3), of a plurality of objects, the detection means being capable of associating with each detected object at least coordinates (x i5 y ; ) as a function of time, the detection system (1) being characterized in that it further comprises: - means of calculating (12) an uncertainty (Ui) for each detected object, as a function of at least the coordinates (x i5 y ; ) depending on the time associated with this object, - means of selecting (14) an object, from among the plurality of objects, according to the uncertainties (U ; ) calculated for these objects by the computing means (12), and - means for modifying (18) the position of at least one sensor (SI, S2, S3), capable of adjusting the position of the selected object in a field of vision of said at least one sensor (SI, S2, S3).

Images