CZ308605B6 - Sensory system with fast group fusion to obtain a model of objects in its vicinity - Google Patents

Abstract

The invention is a fast group fusion sensor system for obtaining a model of objects in its vicinity and is a system for group fusion and for creating an image of objects in the vicinity of the sensor system on which the invention is. The object imaging process is based on an iterative prediction-correction mechanism and includes at least two sensors, an output module, a prediction module, and a correction module, the correction module also includes a group module, a sampling module, and a measurement model module.

Description

Sensory system with fast group fusion to obtain a model of objects in its vicinity

Field of technology

The invention relates to a fast group fusion sensor system for obtaining a model of objects in its vicinity using the fusion of data from a number of sensors to form an image of the environment, for example in vehicles such as autonomous vehicles.

Prior art

At present, technology around autonomous vehicle control and other sensor systems is on the rise. More and more companies are dealing with autonomous control systems, assistance systems, systems for creating the image of objects in the environment sensed by certain sensors both in the automotive industry and in other sectors dealing with other means of transport such as the aerospace industry.

Definitely, an autonomous vehicle, or also a self-driving vehicle, is a vehicle that does not need a driver to operate and is completely oriented with the help of computer systems and sensors that detect the vehicle's surroundings and determine its route, while sensors to detect the vehicle's surroundings, position and speed are more species, such as camera, radar, lidar, etc.

An autonomous vehicle, a vehicle with assistance systems, or a system for creating an image of objects in the environment sensed by certain sensors needs a plurality of sensors to operate, and, in general, the larger the number and total number, the more reliable the measurement result. The individual measurements recorded by the individual sensors are usually not evaluated separately, but these measurements are combined into one complex image of the environment. This achieves a higher reliability of the measurement results and the individual measured values can be verified using measured values from other sensors.

In a simplified sense, the output of individual sensors is the status of objects located in the vicinity of an autonomous vehicle, a vehicle with assistance systems, or another device using a system to create an image of objects in its vicinity. Sometimes individual objects may overlap partially or completely, or are simply too close, which causes problems and requires further monitoring of these objects. There are two basic approaches to dealing with this problem.

The global approach perceives the whole world as one super-object, resp. one large group of objects. This approach has one advantage - it does not exclude any possible interaction between any objects and therefore, the accuracy of the measurement is maximum. The disadvantage of this approach is the computational complexity for combinations of interactions of all objects. Using the global approach, the number of possible combinations of interactions increases exponentially with the increasing number of objects. Therefore, this approach cannot be effectively applied to, for example, autonomous vehicles, vehicles with assistance systems, or other devices using systems to create an image of objects in the vicinity of such a device, because the results cannot be obtained in real time.

The local approach considers all objects to be separate and independent of each other. In this approach, no interactions between individual objects are considered, which significantly reduces the computational complexity. Computational complexity increases linearly with increasing number of objects. The disadvantage and at the same time the reason why this system is not advantageous in this embodiment to be fully applied, for example, to an autonomous vehicle, vehicles with assistance systems, or to other devices using systems to create images of objects in the vicinity of such devices is that neglecting all object interactions useful information and the reliability of the measurement is much lower.

-1 CZ 308605 B6

Another problem is that each sensor has its own specifics, inaccuracies and weaknesses. The camera can capture the shape of the object, not its distance, while the radar has the opposite. Differences in measurements can also be found for sensors of the same type (eg between two cameras). Discrepancies between measurements can also be observed with one sensor, eg the camera provides more accurate measurements for objects in the center of its field of view than for objects at the edges of the field of view. Thus, for example, there are situations where, for example, two sensors detect different distances of one object at the same time, or one sensor detects two significantly different distances of one object at two close times. In this case, it is inappropriate to treat the distance values as absolute values, but it is necessary to evaluate the probability distribution of the distance value. The most commonly used way to solve these problems is through code. Such a solution works in simple situations, however, in complex situations, such as in a city traffic, it is unusable.

US 9963215 B2 describes a solution for fusing data from sensors for autonomous vessels. In this solution, the problem is viewed comprehensively. A probabilistic model is created, on the basis of which the "cluster hypothesis" and the "global hypothesis" are later created, ie groups of objects are created that can interact with each other and these are enumerated together. However, such a solution has a significant disadvantage and that is the fact that a small number of hypotheses are always selected with the highest probability and the others are completely eliminated. Such an approach may work to some extent for seagoing vessels, as water traffic is significantly less dense and more predictable, and thus statistically less likely scenarios that are not anticipated or less likely to apply. For example, such an approach is unsuitable for road traffic because the situation on the road changes every second, in many cases not at all predictable, and therefore it is not appropriate to eliminate less likely situations / locations of objects to maintain the safety of all road users.

Another close solution is disclosed in US 7289906 B2. This solution deals with the fusion of data from sensors using Kalman filters and Gaussian probability distribution. It is the Gaussian distribution that has one significant disadvantage if the measurement is significantly inaccurate / slow. If the sensors detect that there is an object / car in the right lane with a certain probability and with a certain probability in the left lane (in this case there is only a shadow, unusual reflection of light in the left lane, or anything that mimics the presence of the object / car ). According to the Gaussian distribution, the object would be somewhere between these "objects", which will be incorrect in both cases.

Therefore, it would be appropriate to find a solution that would combine the advantages of the global approach, ie high accuracy and reliability of data fusion, and the advantages of the local approach, ie low computational complexity, and thus fast data processing, while eliminating the disadvantages of both approaches. At the same time, the new solution should be able to work with a probability distribution other than Gaussian and would be able to achieve high accuracy and repeatability of measurements in real time.

The essence of the invention

The sensor system with fast group fusion to obtain a model of objects in its vicinity provides a solution to the problems described above. The invention combines the advantages of the already mentioned global and local approach and eliminates their disadvantages. The invention solves the shortcomings of the cited solutions in the cited documents, ie it works with the whole probability distribution of the state of a given object or group of objects, and not only with the most probable ones, so the invention is applicable to applications such as sea or air transport and road applications. operation. Furthermore, the invention can work with a general probability distribution, not just a Gaussian one, and is therefore applicable to a wide range of applications.

- 2 CZ 308605 B6

The invention includes a fast group fusion sensor system for modeling an object in its vicinity, comprising at least two sensors, at least one processor, a prediction module, a group module, a measurement model module, a sampling module, and an output module comprising the steps of: :

- sending the probability distributions of the predicted states of the objects from the prediction module to the group module, where the objects are included in at least one group of objects described by the probability distribution of the state of the group of objects, the distance in the measuring space between every two objects belonging to the same group of objects a given threshold, where the threshold is a preset fixed value, or a floating value depending on the particular measurement;

- sending the states of the objects recorded by the sensors from the sensors to the measurement model module, where they are converted into a probabilistic distribution of the states of the objects recorded by the sensors;

- sending probabilistic state distributions of the objects detected by the sensors and probabilistic state distributions of at least one group of objects to the sampling module;

generating a corrected probability state distribution of the group of objects based on the probability distributions of the states of the objects detected by the sensors, and

- the probability distribution of the state of this group of objects taking place in the sampling module by means of Gibbs sampling;

- selecting the corrected probability distributions of the states of the objects belonging to the group of objects on the basis of the corrected probability distributions of the states of the group of objects to which these objects belong;

- sending corrected probabilistic distributions of the states of the objects to the output module in order to create an image of the objects in the vicinity of the device;

- sending corrected probability distributions of object states to the prediction module, wherein the prediction module includes an object behavior model;

- generating probabilistic distributions of predicted states of objects, based on the model of object behavior and corrected probabilistic distributions of states of objects, wherein the probabilistic distribution of predicted states of objects includes an estimate of the future state of objects

The invention further includes a fast group fusion method for obtaining an object model in the vicinity of a sensor system comprising at least one processor, at least two sensors, a prediction module, a group module, a measurement model module, a sampling module, and an output module. , characterized in that it comprises the steps of:

- sending the probability distributions of the predicted states of the objects from the prediction module to the group module, where the objects are included in at least one group of objects described by the probability distribution of the state of the group of objects, the distance in the measuring space between every two objects belonging to the same group of objects a given threshold, where the threshold is a preset fixed value, or a floating value depending on the particular measurement;

- sending the states of the objects recorded by the sensors from the sensors to the measurement model module, where they are converted into a probabilistic distribution of the states of the objects recorded by the sensors;

-3GB 308605 B6

- sending probabilistic state distributions of the objects detected by the sensors and probabilistic state distributions of at least one group of objects to the sampling module;

- generating a corrected probability state distribution of the group of objects based on the probability distributions of the states of the objects recorded by the sensors and the probability distribution of the state of this group of objects taking place in the sampling module by Gibbs sampling;

- selecting the corrected probability distributions of the states of the objects belonging to the group of objects on the basis of the corrected probability distributions of the states of the group of objects to which these objects belong;

- sending corrected probabilistic distributions of the states of the objects to the output module in order to create an image of the objects in the vicinity of the device;

- sending corrected probability distributions of object states to the prediction module, wherein the prediction module includes an object behavior model;

- generating probabilistic distributions of predicted states of objects, based on the model of object behavior and corrected probabilistic distributions of states of objects, wherein the probabilistic distribution of predicted states of objects includes an estimate of the future state of objects

In an alternative embodiment, the invention includes a method of driving a fast group fusion vehicle to obtain a model of objects in its vicinity comprising at least two sensors, at least one processor, a prediction module, a group module, a measurement model, a sampling module, and an output module. such a device, characterized in that it comprises the steps of:

- sending the probability distributions of the predicted states of the objects from the prediction module to the group module, where the objects are included in at least one group of objects described by the probability distribution of the state of the group of objects, the distance in the measuring space between every two objects belonging to the same group of objects a given threshold, where the threshold is a preset fixed value, or a floating value depending on the particular measurement;

- sending the states of the objects recorded by the sensors from the sensors to the measurement model module, where they are converted into a probabilistic distribution of the states of the objects recorded by the sensors;

- sending probabilistic state distributions of the objects detected by the sensors and probabilistic state distributions of at least one group of objects to the sampling module;

- generating a corrected probability state distribution of the group of objects based on the probability distributions of the states of the objects recorded by the sensors and the probability distribution of the state of this group of objects taking place in the sampling module by Gibbs sampling;

- selecting the corrected probability distributions of the states of the objects belonging to the group of objects on the basis of the corrected probability distributions of the states of the group of objects to which these objects belong;

- sending corrected probabilistic distributions of the states of the objects to the output module in order to create an image of the objects around the car;

-4GB 308605 B6

- sending corrected probability distributions of object states to the prediction module, wherein the prediction module includes an object behavior model;

- generating probability distributions of the predicted states of the objects, based on the object behavior model and the corrected probability distributions of the states of the objects, wherein the probability distribution of the predicted states of the objects comprises estimating the future state of the objects;

- change of at least one of the driving parameters of the car, including the speed or direction of travel, based on the corrected probability distributions of the states of the objects that have been sent and processed in the input module.

Explanation of drawings

The essence of the invention is further elucidated on the basis of examples of its embodiment, which are described with the aid of the accompanying drawings, where in:

Fig. 1 schematically shows the interaction of a prediction module with individual elements of a correction module;

Fig. 2 is a diagram of the whole system;

Fig. 3 shows a vehicle that uses a system to create an image of objects in the environment and a method of driving the vehicle.

Examples of embodiments of the invention

For recording the surroundings of the sensor system, a large number of reliable sensors of 3 different types are inevitable in the first place. In the first embodiment of the invention, visual cameras, RADAR and LiDAR are used as sensors 3. Alternatively, however, two or more cameras may be used, a combination of any number of radars and cameras, a combination of any number of lidars and cameras or any combination of the above and other sensors 3 not disclosed in this application capable of sensing at least one of a group of parameters including speed, size, distance, shape, color and type of object.

Cameras are the most advantageous and most used sensors 3 for sensor systems, such as autonomous vehicles or vehicles with assistance systems. Standard cameras provide two-dimensional images (2D cameras), however, sometimes cameras providing virtual three-dimensional images (3D cameras) are also used. In the case of 3D cameras, four to six cameras are needed to create a three-dimensional image. In a first embodiment of the invention, when applied to a sensor system, five cameras are used, where one camera is pointed in front of the sensor system, one camera is pointed upwards above the sensor system, one pointed backwards, one to the left and one to the right of the sensor system.

LiDAR is a device usable for detecting the spatial monochromatic image of individual objects captured by another sensor 3 (eg a camera). Another advantage of LiDAR is the ability to capture small objects that RADAR could not detect, due to the fact that the wavelength of electromagnetic radiation used in it is much lower than that of radar. The disadvantage compared to RADAR is that it is not usable for long distances. The limit distance for reliable usability of LiDAR usually does not exceed 50 to 200 meters. The last disadvantage is that LiDAR is not usable in cloudy weather or rain. In a first embodiment of the invention, when is

-5GB 308605 B6 applied to the sensor system, four LiDAR sensors 3 are used, one pointing forward to the right of the sensor system, one pointing forward to the left, one pointing backwards to the right, and one pointing backwards to the left.

RADAR is a reliable and affordable sensor 3 used to determine the distance of individual objects, even in bad weather and over very long distances. However, its use is limited by the fact that due to the large wavelength of the electromagnetic radiation used, it is not possible to detect small objects and also to examine their exact shape. In a first embodiment of the invention, when applied to a sensor system, there are six RADAR sensors 3 on the sensor system, two facing forward, one facing right, one facing left and two facing backwards behind the sensor system.

Each spatial formation detected by at least one sensor 3 is called an object. Different sensors 3 record its various parameters, which are most often speed, size, shape, distance, color, or type of object, the term object type meaning whether it is, for example, a human, car, animal, etc. At least one parameter is measured by at least one sensor 3. The value of each parameter measured by sensor 3 is absolute, ie it has a specific value. As already mentioned, the values of one parameter for one object measured by two sensors 3 at the same time may differ, and therefore the system must decide which of these values corresponds more to reality, while the most advantageous and reliable solution to this problem is to convert all values to probabilistic. layout. The summary parameter of each object carrying information about the measured parameters is the state of the object. At least one state 5 of the object detected by the sensors 3 is sent for fusion in order to create an image of the objects in the vicinity of the sensor system. The fusion itself takes place on the basis of an iterative mechanism based on the principle of prediction - correction, while the function of the whole mechanism is shown in Fig. 2. This means that one iteration of fusion can be divided into two basic phases, which are prediction phases and correction phases. It is a mechanism similar to the one by which the human brain processes information about its environment. In the case of the human brain, its sensors 3 are eyes, ears, taste canals, etc., and each signal from sensor 3 is compared with an image of the surrounding world created by the human brain based on previous data and constantly updates this image of the surrounding world.

The fast group fusion sensor system for creating a model of objects in its vicinity comprises at least one processor 11 and at least one storage memory. In a first embodiment of the invention, all the described modules use one common processor 11 and one common storage memory. In an alternative embodiment, each module uses its own storage memory and its own processor 11, or some modules share processors 11 and storage memories with each other.

The correction phase takes place in correction module 2. During the correction phase, the states of the objects are corrected. The three main processes taking place in the correction phase are the conversion of the states of 5 objects detected by the sensors 3 into a probability distribution, the division of objects into groups of 6 objects and Gibbs sampling.

The input information to the correction phase is the states 5 of the objects recorded by the sensors 3, which carry specific values of individual parameters describing the object and the probability distribution 4 of predicted states of objects, the probability distributions 4 of predicted states of objects . Both the states 5 of the objects detected by the sensors 3 and the probability distribution 4 of the predicted states of the objects carry information about at least one of the parameters of the object including speed, distance, size and shape. The states of the 5 objects detected by the sensors may directly contain information about one particular parameter, such as the distance of the object, or they may carry this information indirectly. An example of indirect parameter information is, for example, the case where the distance of an object at time t1 and the distance of an object at time Ϊ2 are known, and these parameters indirectly carry information about the velocity of the object. The speed of this object is then determined in the first embodiment in the module 22 of the measurement model. In the first embodiment

The probability distributions 4 of the predicted object states are sent from the prediction module 1 to the group module 21. The group module 21 is implemented in the correction module 2. The group module 21 includes a storage memory for storing input information and a processor 11 on which the group module is implemented. an algorithm that classifies all objects described by probabilistic distributions of 4 predicted states of objects into at least one group of 6 objects depending on their mutual distance in the measurement space, where any two objects included in different groups of 6 objects are more distant from each other in the measurement space, than the given threshold. Any two objects included in the same group of 6 objects are spaced apart in the measurement space less than the given threshold. In the first embodiment, the threshold is a preset distance value in the measuring space. In an alternative embodiment, the threshold value is movable, the so-called floating value, which means that the value of the distance in the measuring space corresponding to the threshold changes dynamically during iterations. The minimum number of groups of 6 objects is one and the maximum number of groups of 6 objects is not limited. The current number of groups of 6 objects is always defined by individual objects and their mutual distances in the measurement space. It is true that the minimum number of objects in a group of 6 objects is two and one object can belong to only one group of 6 objects. A group algorithm is any algorithm whose output meets the above requirements, such as the following:

In the first step, one group of objects is created for each object, which contains this object;

In the second step, if there are two groups of objects s _o ias ₀ 2, some two objects oi belong to Soi and 02 belong to s ₀ 2 and at the same time oi and 02 are spaced apart from each other in the measurement space less than or equal to threshold, groups of objects with _o ias ₀ 2 are merged into one group;

The second step is repeated until there are at least two groups to merge, otherwise the process of dividing into groups of 6 objects is completed.

The output of the group module 21 is at least one group 6 of objects, which is described by the probabilistic distribution of the state of the group of 6 objects, the group of 6 objects containing objects whose mutual distances in the measurement space are less than a defined threshold. Each group of 6 objects consists of at least two objects. Each object belonging to a group of 6 objects is characterized by a probabilistic distribution of the state of the object. The probability state distribution of each object belonging to the group of 6 objects contains information about at least one of the parameters such as speed, distance, shape, size, color and type of the object. The probability state distribution of a group of 6 objects contains information about at least one of the parameters such as speed, distance, shape, size, color and type of the object.

In the case where the probability distributions 4 of the predicted states of objects are not available, such as the first iteration, the probability distributions of the 4 predicted states of the objects are not sent to the group module 21 and the group module 21 remains inactive within such iteration.

The states of the 5 objects detected by the sensors are sent from the sensors 3 to the module 22 of the measurement model. The states of 5 objects recorded by the sensors are described by specific measured values for at least one of parameters such as speed, size, shape, distance, color, object type, etc. The measurement model module 22 includes temporary storage memory and processor 11 and is adapted to convert specific parameter values contained. in the states of 5 objects recorded by sensors in the probability distribution, each object is subsequently described by a complex parameter, which is the probability distribution 10 of the state of the object recorded by sensors, and this complex parameter includes combinations of all parameters measured for the object by sensors 3. the measurement is thus the probability distribution of 10 states of objects detected by the sensors.

-7 CZ 308605 B6

The states of the 5 objects detected by the sensors are sent from the sensors 3 to the module 22 of the measurement model. The measurement model module 22 comprises a temporary storage memory and a processor 11 and is adapted to convert specific values of parameters contained in the states 5 of objects recorded by the sensors and stored measurements recorded by the sensors 3 to a probability corresponding to the probability that the sensors 3 would recorded by the sensors recorded the measurement. This probability and its distribution is described by the probability distribution of 10 states of objects detected by sensors. The measurement model module 22 is repeatedly used by the Gibbs sampler in the sampling module 23.

In the first exemplary embodiment, the probability state distribution of the at least one group 6 of objects together with the probability distributions 10 of the object states recorded by the sensors is sent to the sampling module 23. The sampling module 23 comprises temporary storage and a processor 11 on which a Gibbs sampler is implemented. By means of a Gibbs sampler, the probability distributions 10 of the states of the objects recorded by the sensors and the probability distributions of the states of at least one group of 6 objects are corrected in the sampling module 23. In the first phase of the correction process, the so-called corrected probability state distribution of a group of 6 objects is calculated based on the probability state distribution of the same group of 6 objects and the probability distributions of 10 states of objects detected by the sensors. The corrected probabilistic distribution of the state of a group of 6 objects describes the parameters of this group of 6 objects as a whole and at the same time the parameters of individual objects contained in this group of 6 objects.

When calculating the corrected probability distribution of the state of a group of 6 objects, improbable states of a group of 6 objects are not completely eliminated. To simplify the computational complexity of the whole process, the number of individual possible states of a group of 6 objects within its probability distribution is reduced to a predefined number of states, while in the first embodiment this defined number of states is 100,000. The reduction has the same shape, but is composed of a smaller number of individual possible states, which achieves the preservation of measurement accuracy and at the same time simplifies computational complexity. And this is achieved by using Gibbs sampling, which takes into account each, even the least probable state of an object included in the probability distribution of the state of a group of 6 objects. Each state of a group of 6 objects included in the probability distribution of the state of this group of 6 objects will affect the corrected probability distribution of the state of the group of 6 objects, however, the degree of such influence is determined by the probability of that state. This means that less probable states from the probabilistic state distribution of a group of 6 objects will affect the corrected probability state distribution of a group of 6 objects less than more probable states from the probabilistic state distribution of a group of 6 objects and the possibility of applying less probable states, ie the possibility that less probable states are the truth is given precisely by their probability. The Gibbs sampling process is described as follows:

There is (from the previous procedure) a group of 6 n objects that need to be evaluated in the system together. The probability distribution of the state of each object i is represented by the set of samples S, (i z

The aim is to obtain a common probability distribution for a group of objects, ie p samples. The Gibbs sampling process uses the evaluation function f (X), where Xj = (xij, X2j, ... x _n] ), which sends specific states of X objects to the measurement model module 22 and returns the probability calculated by the measurement model module 22

The first sample is randomly selected - ie Xo = (xio, X20 ... x _n o)

For j = 1 to p

For k = 1 to n

-8GB 308605 B6

Randomly, another sample is selected from the set of samples Sk, denote it x'y and create an alternative common state X'j = (xij, xy, ... x'kj, ... x _nj ) by replacing Xkj with x'kj;

It is decided to accept a new sample:

a. random alpha from uniform distribution is selected (0..1)

b. if f (X’j) / f (Xj) <= alpha accepts x’kj and thus Xj: = Xy end k;

sample Xj is added to the resulting common distribution;

endj;

The result is p samples from a common distribution;

At the same time, the individual components of these samples represent a new probability distribution for individual objects after the application of a given measurement.

In the second phase of the correction in the sampling module 23, the corrected probability distribution 7 of the state of the objects contained in the group 6 of objects is determined, for which the corrected probability distribution of the state of the group 6 of objects has been calculated. Determining the corrected probability state distribution 7 of objects contained in the group 6 of objects for which the corrected probability state distribution of the group 6 objects has been calculated is enabled by the fact that the corrected probability state distribution of the group 6 objects directly contains information on the probability distribution of object states in this group 6. objects contained, so it is a matter of selecting this information.

The corrected probability distribution 7 of the object state describes at least one object parameter such as speed, distance, size, shape, color and type of the object and determines the corrected state of the object. The corrected probability distribution 7 of the state of the object is then sent to the output module 8. Likewise, this corrected probability distribution 7 of the state of the objects is sent to the prediction module L

The output module 8 combines the corrected probability distribution 7 of the states of the objects belonging to the groups 6 of objects and forms from them an image of the objects in the vicinity of the sensor system. The output module 8 can be, for example:

- a vehicle control system which, on the basis of such an image of objects around the autonomous vehicle, changes at least one of the car's driving parameters including direction of travel, speed, power unit output, brake system settings, wheel rotation settings, gear changes or signaling changes. A vehicle using this system is shown in Figure 3;

- a vehicle assistance system that alerts the driver / pilot to possible obstacles or dangers;

- a system for the projection of a complex four-dimensional map of a bounded space.

System for monitoring free parking spaces in the car park, where various sensors 3 are placed in the car park monitoring the parking space occupancy with the subsequent possible function of setting the information board free / occupied, and / or setting navigation boards directing the driver to free parking spaces.

In the prediction module 1, the prediction phase of the iteration takes place. A detail on the prediction phase of the iteration is shown in Fig. 1. The prediction module 1 comprises a storage memory 9 of the prediction module intended for storing probability distributions 4 of predicted states of objects and a processor 11 with an implemented object behavior model. Corrected probability distribution of 7 states of objects from iteration,

-9EN 308605 B6 which has been described above, after sending to the prediction module 1, it becomes a probabilistic distribution of the predicted states of the objects for the next iteration.

Examples of prediction using the object behavior model:

The monitored object is auto, the parameters of the state are: position (x, y) and speed (y _x , in _y ), so the whole state is (x, y, in _x , you).

The prediction has inputs: 1. corrected probability distribution 7 of the state of the object in time that is n samples in the form (x, y, v _x , v _y ), 2. elapsed time for prediction: dt.

The output is the probability distribution 4 of the predicted state of the object in time to + dt, which is m of samples in the form (x, y, v _x , v _y ).

Example

The input is 100 samples (0 + - 2, 0 + - 2, 10, 0) so the car is approximately in position (0, 0) plus or minus 2 meters, and travels at a speed of 10 m / s forward, dt = 1 s, so prediction 1 s ahead.

The output is 100 samples (10 + -3, 0 + -2.5, 10 + - 1, 0 + - 1) so it is assumed that car 1. will be in position (10, 0) plus or minus (3, 2.5) - uncertainty increases, 2. travels at speed (10, 0) plus or minus 1 - speed uncertainty increases.

The prediction algorithm uses:

- physical laws - movement of a car at a given speed;

- knowledge of the likely behavior of the car / driver, such as not likely to turn too hard.

As a result of using the probability distributions of 4 predicted object states in the correction phase of the next iteration, a system using a sensor system with fast group fusion of data to obtain a model of objects in its vicinity behaves intelligently. That is, if the probability distribution 10 of the state of the object detected by the sensor in the next iteration changes radically compared to the probability distribution 4 of the predicted state of the same object, the predicted state of the object being affected by all past corrected object states, e.g. due to an error on one sensor 3. will react flexibly, which means that it registers the deviation by the corrected probability distribution 7 of the state of the object of the next iteration including information on the probability distribution of the state of the object from the erroneous re-measurement from sensor 3, thus also information on the probability distribution of 10 states of objects detected by the sensors; this corrected probability distribution 7 of the state of the object becomes the probability distribution 4 of the predicted state of this object in the following iteration.

Claims (9)

PATENT CLAIMS A fast group fusion method for obtaining a model of objects in the vicinity of a sensor system comprising at least one processor (11), at least two sensors (3), a prediction module (1), a group module (21), a measurement model module (22), a sampling module (23) and an output module (8), wherein the fast group fusion method comprises the steps of: - sending probabilistic distributions (4) of the predicted states of objects from the prediction module (1) to a group module (21), where the objects are included in at least one group (6) of objects, which is described by a probabilistic distribution of the state of the group (6) of objects, in the measuring space between every two objects belonging to the same group of objects is less than a given threshold, where the threshold is a preset fixed value, or a floating value depending on the particular measurement; - sending the states (5) of the objects recorded by the sensors from the sensors (3) to the module (22) of the measurement model, where they are converted into a probability distribution (10) of the states of the objects recorded by the sensors; - sending probability distributions (10) of the states of the objects recorded by the sensors and probabilistic distributions of the states of the at least one group (6) of objects to the sampling module (23); - generating a corrected probability state distribution of the group (6) of objects based on the probability distributions (10) of the states of the objects recorded by the sensors and the probability state distribution of this group (10) of objects taking place in the sampling module (23) by Gibbs sampling; - selecting the corrected probability distributions (7) of the states of the objects belonging to the group (6) of objects on the basis of the corrected probability distributions of the states of the group (6) of objects to which these objects belong; - sending corrected probability distributions (7) of the states of the objects to the output module (8) in order to create an image of the objects in the vicinity of the device; - sending corrected probability distributions (7) of the states of the objects to the prediction module (1), the prediction module (1) comprising a model of the behavior of the object; - generating probability distributions (4) of the predicted states of the objects, based on the object behavior model and the corrected probability distributions (7) of the states of the objects, the probability distribution (7) of the predicted states of the objects comprising estimating the future state of the objects. The fast group fusion method according to claim 1, characterized in that the states (5) of the objects detected by the sensors comprise at least one of a group of parameters including speed, size, distance, shape, color, and object type. Fast group fusion method according to claim 1, characterized in that the corrected probability distribution (7) of the state of the object which was generated in the sampling module (23) in one iteration is sent to the prediction module (1), where it becomes a probability distribution (4) the predicted state of the same object for the next iteration. Fast group fusion method according to claim 2, characterized in that when generating a corrected probability state distribution of a group (6) of objects in the sampling module (23) based on probability distributions (10) of object states recorded by sensors and probability state distribution of the same group (6) objects is corrected -11 CZ 308605 B6 probabilistic state distribution of a group of objects (6) affected by each state from the probabilistic distribution of a group (6) of objects, while less probable states from the probabilistic distribution of a group (6) of objects have less influence than more probable states from the probabilistic distribution of a group (6). ) objects. A method of driving a fast group fusion vehicle to obtain a model of objects in its vicinity comprising at least two sensors (3), at least one processor (11), a prediction module (1), a group module (21), a measurement model (22), sampling module (23) and an output module (8), wherein the rapid group fusion of data in the vicinity of such a vehicle is characterized in that it comprises the steps of: - sending probabilistic distributions (4) of the predicted states of objects from the prediction module (1) to a group module (21), where the objects are included in at least one group (6) of objects, which is described by a probabilistic distribution of the state of the group (6) of objects, in the measuring space between every two objects belonging to the same group (6) of objects is smaller than a given threshold, where the threshold is a preset fixed value, or a floating value depending on the particular measurement; - sending the states (5) of the objects recorded by the sensors from the sensors (3) to the module (22) of the measurement model, where they are converted into a probability distribution (10) of the states of the objects recorded by the sensors; - sending probability distributions (10) of the states of the objects recorded by the sensors and probabilistic distributions of the states of the at least one group (6) of objects to the sampling module (23); - generating a corrected probability state distribution of the group (6) of objects based on the probability distributions (10) of the states of the objects recorded by the sensors and the probability state distribution of this group (6) of objects taking place in the sampling module (23) by Gibbs sampling; - selecting the corrected probability distributions (7) of the states of the objects belonging to the group (6) of objects on the basis of the corrected probability distributions of the states of the group (6) of objects to which these objects belong; - sending corrected probability distributions (7) of the states of the objects to the output module (8) in order to create an image of the objects around the car; - sending corrected probability distributions (7) of the states of the objects to the prediction module (1), the prediction module (1) comprising a model of the behavior of the object; - generating probability distributions (4) of the predicted states of the objects, based on the object behavior model and the corrected probability distributions (4) of the states of the objects, the probability distributions (4) of the predicted states of the objects comprising estimating the future state of the objects; - change of at least one of the group of driving parameters of the car including the speed and direction of travel based on the corrected probability distributions (4) of the states of the objects that have been sent and processed in the input module. Vehicle control method according to claim 5, characterized in that the states (5) of the objects detected by the sensors comprise at least one of a group of parameters comprising speed, size, distance, shape, color and type of object. Vehicle control method according to claim 5, characterized in that the corrected probability distribution (7) of the state of the object which was generated in the sampling module (23) in one iteration is -12 CZ 308605 B6 sent to the prediction module (1), where it becomes a probability distribution (4) of the predicted state of the same object for the next iteration. Vehicle control method according to claim 5, characterized in that when generating a corrected probability state distribution of a group (6) of objects in the sampling module (23) based on the probability distributions (10) of object states recorded by sensors and the probability state distribution of the same group ( 6) objects, the corrected probability distribution of the state of the group of objects (6) is affected by each state from the probability distribution of the group (6) of objects, while less probable states from the probability distribution of the group of objects (6) have less influence than more probable states from the probabilistic distribution of the group (6). ) objects. A fast group fusion sensor system for obtaining a model of objects in its vicinity implemented on at least one processor (11) based on an iterative prediction-correction mechanism comprising: - at least two sensors (3) adapted to monitor the objects and send the states (5) of the objects recorded by the sensors to the module (22) of the measurement model; - a prediction module (1) with a built-in storage memory (9) of the prediction module for probabilistic distribution (4) of predicted states of objects; - a correction module (2), which comprises a group module (21), a measurement model module (22) and a sampling module (23); - a group module (21) which is adapted to receive probabilistic distributions (4) of predicted states from objects from the prediction module (1) and sort objects into at least one group (6) of objects, the state being described by probabilistic state distributions of a group (6) of objects and sending a probabilistic state distribution of the at least one group (6) of objects to the sampling module (23); - a measurement model module (22) adapted to receive the states (5) of the objects recorded by the sensors from the sensors and convert the states (5) of the objects recorded by the sensors into a probability distribution (10) of the states of the objects recorded by the sensors and send them to the sampling module (23) ; - a sampling module (23) comprising a Gibbs sampler, the sampling module (23) being adapted to receive probability state distributions of groups of objects (6) and probability distributions (10) of object states detected by sensors and further adapted to correct probabilistic state distributions of at least one group (6) objects and probability distributions (10) of object states recorded by sensors using a Gibbs sampler to produce a corrected probability state distribution of at least one group (6) of objects and selecting a corrected probability distribution (7) of object states contained in the corrected group state probability distribution (6) objects to which these objects belonged and sending a corrected probability distribution (7) of the state of the at least one object to the output module (8) and the prediction module (1); - an output module (8) which is adapted to receive a corrected probability distribution (7) of the states of the objects.