The technical field relates generally to methods and devices for determining kinematics of a target, especially a vehicle, and may be used in autonomous driving systems.
For autonomous or semi-autonomous driving systems, one important factor for radar systems is to determine correctly and accurately the kinematics data—meaning position, speed, and heading—of a moving object set as a target. As radar sensors have a fair amount of noise in the detection of surrounding world, smoothing algorithms (Kalman-filter, polynomial regression, low-pass filter) are used to calculate the target object's parameters. These algorithms reduce the noise coming from the sensor, but the accuracy is never 100%, as a standard deviation always exists.
Nevertheless, when testing the blind spot detection functionality on its boundaries, for example, it is important to have an almost 100% accuracy—specific test scenarios allow only extremely small deviations (see “Blind Spot Detection System Confirmation Test for NCAP—Memorandum/Report—NHTSA—2015-0119-0025”). The test is done by comparing the data coming from a high precision GPS with objects which are exactly at the blind spot detection zone limit and even deviations with 0.1 s of late warnings are considered failures.
On the other side, the current radar hardware does not have such a great precision as a GPS module, and the noise coming from the radar must be reduced. Two degrees deviation in angle may exist in the best zone, this means 0.34 m deviation at 10 m. The noise can be reduced by Kalman filter to a certain degree, but not totally.
Furthermore, the standard deviation of an absolute heading calculation is even bigger than for position and speed, as it is harder to reduce the noise. If a stricter Kalman filter is used, then the maneuvers of the target object are missed entirely. For example, the absolute heading of a vehicle cannot change more than five degrees from one cycle to another. Moreover, the absolute heading of a vehicle is strictly constant/increases/decreases through a maneuver, it has only one type of change or no change at all. The same goes for acceleration, there is no fluctuation in real world such as in one cycle there is acceleration, in the next deceleration. Still, these fluctuations exist in radar measurement due to the noise the measurement brings, the smaller the noise the smaller the fluctuation, or the measurement is not real. Anyway, the change of parameters as position, heading, speed, acceleration is always simple in normal case scenarios and can be mapped by a set of linear function defined in a Markov-chain.
In this respect, the Korean patent KR101402206B1 discloses a multiple target tracking method, more particularly, a multiple target tracking method using distance information of a target, feature information-based Kalman filter, and joint probability data-association method. According to that document, a joint probability data-association method generates not only normalized distance squared distance information but also an effective matrix through dual conditions based on the correlation information of an image. By using this, the identification of other objects, which closely exist, can be possible and the complexity decreases as the number of cases which are taken into account decreases.
Another patent document, U.S. Pat. No. 5,633,642 A, discloses a radar method and a corresponding device. The method implies estimating an azimuth angle of each target object from determined variables of distance or range, relative speed and relative acceleration after Kalman filtering and separating out target objects having a physically impossible behavior (tracking and prediction). That is used to determine which target objects are located on a road-way occupied by a vehicle and which are the most dangerous thereof.
Nevertheless, using only smoothing algorithms to reduce the noise, predict and calculate the kinematics of a target object is not enough. In essence, none of the afore-mentioned inventions does not reflect how the objects are moving in the real world.
As such, it is desirable to present a system and method for highly increased accuracy on mapping the movement of target objects. In addition, other desirable features and characteristics will become apparent from the subsequent summary and detailed description, and the appended claims, taken in conjunction with the accompanying drawings and this background.
According to a first aspect, the disclosure provides a method of determining a kinematic of a target in relation to an ego vehicle. The method includes selecting the target when it is detected in a predetermined sensing zone. The method also includes generating a trace of the target based on position, heading, speed and acceleration. The method further includes predicting a curve the target is taking and calculating a radius of the curve using the generated trace. The method also includes projecting n paths possible to be followed by the target, expressed as abstract movement functions of a polynomial degree p and using as input the calculated radius and, after having a trace of at least 2p−1 historical points, adding at each further cycle higher probabilities to the paths projected in a previous measuring cycle which are equal with the newly projected ones, and keeping the projected paths, if the radius remains the same as in the previous measuring cycle. The method further includes computing the current kinematics values by smoothing filter, as predicted kinematics values. The method also includes comparing the predicted kinematics values with the kinematics values resulted from the projected paths and determining the final kinematics values.
According to a second aspect, the disclosure provides a device of determining a kinematic of a target. The device includes an interface and an evaluation unit. The interface is adapted to receive sensor data related to the target. The evaluation unit is adapted to: predict a curve the target is taking and calculate a radius of the curve using the generated trace; project n paths possible to be followed by the target, expressed as abstract movement functions of a polynomial degree p and using as input the calculated radius, after having a trace of at least 2p−1 historical points; at each further cycle, add higher probabilities to the paths projected in a previous measuring cycle which are equal with the newly projected ones and keep the projected paths, if the radius remains the same as in the previous measuring cycle; compute the current kinematics values using an smoothing algorithm, as predicted kinematics values; compare the predicted kinematics values with the kinematics values resulted from the projected paths to determine the final kinematics data values.
According to a third aspect, the disclosure provides a vehicle equipped with a sensor system adapted to provide sensor data of surrounding of the ego vehicle, and a driving assistance system, able to receive signals from a device of determining a kinematic of a target.
According to one embodiment, a most probable path of a target vehicle is accurately chosen after comparing a predicted path with a projected path. Having this projected path, more precise kinematics values are computed for each of the two implied vehicles, so the method performs data validation in real-time. By doing this, almost all the noise coming from sensors can be eliminated, at least in usual, normal situations. Furthermore, as those projected paths are abstract models which reflect reality, further positions and movement of target can be predicted with higher precision.
The method also contributes to a better performance for advanced driving functions, like blind spot detection, lane change assist, front/rear traffic cross alert and so on.
According to an exemplary embodiment of the method, the step of determining the final kinematics values includes the condition that, if the standard deviation of the kinematics values associated to the projected paths is close to the standard deviation of the kinematics values associated to the predicted paths, then the kinematic parameters are those associated to the projected paths, otherwise the kinematics values are those associated with the predicted paths.
According to a particular embodiment of the method, the abstract movement functions are defined as Markov-chain matrices.
According to another embodiment of the method, the radius of the curve taken by the target is obtained by polynomial regression, using the generated trace.
According to another particular embodiment of the method, the smoothing filter used to compute the kinematics values associated to predicted paths is a Kalman filter.
According to another particular embodiment of the method, the smoothing filter used to compute the kinematics values associated to predicted paths is a low-pass filter.
According to an exemplary embodiment of the device, the evaluation unit is adapted to define the abstract movement functions as Markov-chain matrices.
According to another embodiment of the device, the evaluation unit is adapted to output the determined kinematics data to a driving assistance system, in order to assist functions like blind spot detection, lane change support, front and rear traffic cross alert.
Other advantages of the disclosed subject matter will be readily appreciated, as the same becomes better understood by reference to the following detailed description when considered in connection with the accompanying drawings wherein:
A schematic view of an ego vehicle E equipped with such a device 1 and with a sensing system 2, as well as a driving assistance system 5 is illustrated in
The device 1 comprises an interface 4 which is adapted to receive input data from any active form of remote sensing, such as radar, lidar, optic, ultrasonic or infrared sensors, for example, as well as GPS sensors. The received input data are related to a target external to said ego vehicle E, the target T being present in a predetermined scanning zone of the sensing system 2. In the illustrated examples, the traffic participant designated as target T in this description is a vehicle whose direction (angle) of movement over ground has been detected and history of measurement positions has been generated as a trace.
The device 1 includes an evaluation unit 3 adapted to use input data extracted from the trace of path and positions, and to further determine a kinematic of target T according to the method which is described in detail in the following section.
In order to illustrate the way target vehicle T is perceived by sensing systems 2,
It is of interest, for the purpose of this disclosure, to discuss the cases requiring to map the kinematic parameters to a set of functions which reflect the reality. In reality there is noise in the motion of a vehicle; while driving, in 99% of the cases, the usual situation is that only simple movements are made, having a constant acceleration and a constant direction change. If these movements are expressed by mathematical functions, the exact movement of a target vehicle can be predicted and analyzed.
The concept consists into projecting at each cycle a set of paths expressed by movement functions of a polynomial degree n (which describe every kinematic parameter and make a Markov-chain model) from the trace and current parameters of target T and choosing the movement function which defines most accurately the actual path of the target, by comparing the projected paths with a path predicted by using existing smoothing algorithms.
In case none of the projected paths match with the predicted ones, the updated kinematics values from the smoothing algorithms will represent an unusual situation and thus the output data will be the one resulting from the smoothing algorithm. In the end, the goal is to reach a simplified solution where, using a minimal set of parameters, all the other variables/kinematics values to be calculated.
The set of functions which define the Markov-chain models associated to the kinematic of a vehicle (heading, acceleration, speed and position) comes as follow for a certain maneuver, and period of time t:
where Φ(t), aabs(t) are independent functions describing heading and acceleration in absolute coordinates, and
Next comes the simplified model for acceleration, describing a normal behavior, considering how an acceleration pedal is pressed over time, and how the pedal movement modifies the speed of the vehicle—almost linear changes.
The set of functions serialized, to be used in case of pass-by-pass/frame-to-frame computation looks like that:
For every type of maneuver only the definition of the absolute heading functions differs. This heading function is only a simple first degree polynomial. It is based on observation and mapping of how a normal maneuver is done by a vehicle. For a better understanding of how this normal maneuvers are mathematically described,
Functions F1 and F5 describing constant turning to left/right are expressed by the following Markov-chain:
Functions F2 and F4 describing lane changing are expressed by the following Markov-chain:
Function 3 describing straight heading is described by the following equation:
Φ(t)=0 (7).
An exemplary embodiment of the method of determining a kinematic of a target, applied for highway driving scenarios, is represented in
Target vehicle T is considered as having constant speed behind the ego vehicle E, the same speed as ego vehicle E. In this case, there are absolutely no fluctuations of kinematic parameters from one cycle to another, even the position of target vehicle T relative to ego vehicle E remains the same. The radius as extracted from the trace is calculated as being almost zero. By using polynomial regression, speed and acceleration are determined as constant as well, so there is no acceleration, only constant speed. A number of n paths which could be followed by target vehicle T as normal maneuvers are created as abstract movement functions (in this case two, namely F1—heading straight ahead and F2—lane changing), and the associated kinematics values are computed. Also, current values of kinematics are computed using the existing, normal algorithms (Kalman filter, low pass filter etc.) and compared with the kinematics values associated to the respective functions (projected paths). In the end, the chosen path will be the one of constant absolute heading, since it have the highest probability to be followed by target vehicle T. In this manner, speed, position, heading, acceleration is computed 100% correctly and accurately, eliminating all the noise and, most important, no warning is triggered.
Another exemplary embodiment of how the method of determining the kinematic of a target vehicle T, according to the invention, applies for a specific road scenario is shown in
In this case, in the first cycle, as the Kalman filter predicts target vehicle T is going straight, the straight path is chosen as projected path; however, starting with the next cycle the lane changing maneuver is detected, and from the existing data the radius of the followed curve is determined. For the next few cycles target vehicle T is exactly on the predicted path which is the lane changing path (expressed as F2, as it is more important than the combination of left/right curve path expressing the overtaking maneuver), and as soon as target vehicle T finishes the maneuver, the path is computed correctly on the next lane, according to the method.
While certain embodiments of the present invention have been described in detail, those familiar with the art to which this invention relates will recognize various alternative designs and embodiments for practicing the invention as defined by the following claims.
Number | Date | Country | Kind |
---|---|---|---|
17465565 | Oct 2017 | EP | regional |
This application is a continuation of International application No. PCT/EP2018/073947, filed Sep. 6, 2018, which claims priority to European patent application No. 17465565.4, filed Oct. 26, 2017, each of which is hereby incorporated by reference.
Number | Name | Date | Kind |
---|---|---|---|
7102496 | Ernst, Jr. | Sep 2006 | B1 |
20030225519 | Miyahara | Dec 2003 | A1 |
20100017180 | Randler | Jan 2010 | A1 |
20100253493 | Szczerba | Oct 2010 | A1 |
20130231820 | Solyom | Sep 2013 | A1 |
Number | Date | Country |
---|---|---|
104269070 | May 2017 | CN |
102016206550 | Oct 2017 | DE |
WO-0221156 | Mar 2002 | WO |
Entry |
---|
Machine translation of CN-104269070-B (Year: 2017). |
X. Rong, Li. et al. Survey of Maneuvering Target Tracking. Part 1: Dynamic Models, IEEE Transaction on Aerospace and Electronic Systems, Oct. 2003, pp. 1333-1364, vol. 39, No. 4. |
European Search Report dated May 3, 2018 for corresponding European Patent Application No. 17465565.4. |
Number | Date | Country | |
---|---|---|---|
20200255006 A1 | Aug 2020 | US |
Number | Date | Country | |
---|---|---|---|
Parent | PCT/EP2018/073947 | Sep 2018 | US |
Child | 16858497 | US |