• 专利附录
  • 专利说明
  • 权利要求
  • 类似技术
  • 同族专利
  • 引用文献
  • 法律状态
  • 优先权
    • 专利摘要:
    A method of analyzing a cell-cut dynamic scene in which a probability of occupancy of a cell and the probability (s) of movement of the cell are determined by solving the equation comprising determining the speeds and positions of the fictitious particles in the grid as a function of those determined at the (k-1) th iteration and the probability P (V | V-1); the determination of the particles located in each cell according to the determined positions and in that the resolution of the equation, for a cell, is split into the resolution of a static part corresponding to P (O = empty, V = 0 | ZC) and P (O = occupied, V = 0 | ZC) and in the resolution of a dynamic part corresponding to P (O = occ, V = vki, | ZC), i = 1 to nk, where nk is the number of particles determined in cell C for the kth iteration.
    公开号:FR3022049A1
    申请号:FR1455183
    申请日:2014-06-06
    公开日:2015-12-11
    发明作者:Christian Laugier；Amaury Negre；Mathias Perrollaz；Lukas Rummelhard
    申请人:Centre National de la Recherche Scientifique CNRS；Commissariat a lEnergie Atomique CEA；Institut National de Recherche en Informatique et en Automatique INRIA；Commissariat a lEnergie Atomique et aux Energies Alternatives CEA；
    IPC主号:
    专利说明:

    [0001] The present invention relates to a method of analyzing a dynamic scene observed with the aid of a sensor block (s) comprising a dynamic scene, an analysis module and associated computer program. step of defining a grid cut into cells and corresponding to the observed scene; and iteratively computer-implemented steps of: collecting at least one new observation of the sensor block (s) at the kth iteration and determining, based on the new observation collected, a first probability of occupation of each cell of the grid modeling the operation of the sensor block (s); determining, for each cell, at the kth iteration, a second probability of occupation of the cell and a set of probability (s) of movement of the content of the cell as a function of the first probability of occupation of the cell determined at the kth iteration. According to Bayesian perception (BOF) algorithms, the observed scene is represented by a grid of occupation subdivided into cells and is analyzed in terms of occupied cells and their evolution over time. For each cell of this occupation grid and at each iteration of the algorithm, a probability of occupation of the cell is calculated, as well as a distribution of probabilities of movement of the contents of the cell. This distribution is represented in the form of a motion distribution grid, also called neighborhood transition histogram. In FIG. 1, a view of an occupancy grid of the environment G 'is represented of 16 cells out of 16 cells. In this representation, the closer the hatches covering a cell, the greater the value of the probability of occupancy of the cell. A histogram of transition Hist (C) determined for a cell, referenced C, of the grid Oc, is also represented in FIG. 1. Each cell of the histogram determined for the cell C represents, for the time step under consideration, the transition from cell C to said box; each cell of the histogram is thus associated with a distinct velocity value vector; in this case, the cell in the center of the histogram corresponds to the transition from cell C to itself. These algorithms have the advantage of providing a reliable and synthetic analysis of a dynamic scene, but have a number of limitations.
    [0002] In the case of a moving grid, for example corresponding to the environment analyzed from a moving vehicle, spatio-temporal aliasing problems arise. In addition, it is necessary to use occupancy grids and high resolution motion histograms to detect slow movements and small objects, and also to evaluate the speed of very dynamic objects (eg other vehicles). moving). Typically, if the standard of the maximum speed of the vehicle is Vmax, the size of the motion distribution grid maintained during the algorithm must make it possible to represent speeds of -2V'x to + 2Vmax: each iteration of the The algorithm therefore involves estimating the probability of distribution of the velocities in each cell of this grid for each cell. The necessary calculation volume is therefore very important. This problem is crucial, especially in the case of an analysis embedded in a vehicle. For this purpose, according to a first aspect, the invention proposes a method of analyzing a dynamic scene observed using a sensor block (s) of the aforementioned type, characterized in that the second probability of occupancy of the cell and the set of probability (s) of motion of the contents of the cell according to the first probability of occupation of the cell determined at the kth iteration is determined by solving the equation 0 -iv- iP (CAD 0-11 (V -1Z) P (011 IZC) EA00-ivy-i P (CAO -11717 -1Z) where C: identifier of the cell in question; A: identifier of the cell which contained, at the (k-1) th iteration what contains the cell considered at the kth iteration, O: state of occupation of the cell considered, among the states "empty" and "busy"; 0-1: state of occupation of the cell A at the (k-1) th iteration, V: velocity of the cell C considered, V-1: velocity of the cell A at the (k-1) th iteration; : sensor observations from the first iteration to the kth iteration; the speed and the respective position of a set of fictitious particles in the grid are determined at the kth iteration as a function of the velocities, the positions of the particles determined at the (k-1) th iteration and the probability P (V1V-1 ); the method comprises a step of determining the localized particles in each cell according to the determined positions and in that the resolution of the equation, for a cell, is split into the resolution of a static part corresponding to P (0 = empty, V = 01ZC) and P (0 = occupied, V = 01ZC) and in the resolution of a dynamic part corresponding to P (0 = occ, V = i = 1 to nk, where nk is the number of particles determined in the cell C for the kth iteration, the static part of the cell C at the kth iteration being determined as a function of the static part of the cell C determined at the (k-1) th iteration and P (010-1) the probability of P (O = occupied, V = vk, 1ZC) of the dynamic part of the cell C being determined at the kth iteration as a function of the probability P (O = occupied, V = -1) -1,1ZA ) calculated at the (k-1) th iteration for the dynamic part of cell A and P (010-1), where the particle p, determined in the cell lule C at the kth iteration with a velocity Vk was in cell A at the (k-1) th iteration with a velocity -1) -1. The method according to the invention makes it possible to significantly reduce the computing and memory resources required for the determination of the occupation of the cells and the distribution of probabilities of movement during the iterations. In embodiments, the analysis method according to the invention also comprises one or more of the following characteristics: at the end of the kth iteration, one or more pairs (p ,, (-1))), where xk is the position of the particle p, at the kth iteration, are duplicated within a cell (C) or deleted so that the number of pairs per cell is a function of the dynamic part determined for the cell; the method comprises a step of selecting the pair to be duplicated / suppressed, the selection of said pair being performed as a function of the probability P (O = occupied, V = -1) 1ZC) determined at the kth iteration, where vk is the component torque speed; the total number of particles in the grid is constant over the iterations; P (O = empty, V = 01ZC) determined at the kth iteration and noted Pk (0 = empty, V = 01ZC), is determined as a function of the product of a first term, a function of the first probability of occupation and of a second term, said second term being a function of: - Pk-1 (0 = OCC, V = 01ZO). (1 -c) Pk-1 (0 = empty, V = 01ZC) .beta., and / or - coeff (k). pk-1 (0 = occupied, V = ik, 1ZA). (1 - where coeff (-1) is a decreasing function of Fkll, and / or - a probability of occurrence pa of a new object in the observed scene According to a second aspect, the present invention proposes a computer program to be installed in an analysis module of a dynamic scene observed with the aid of a sensor block (s) comprising a step of defining a grid cut into cells and corresponding to the observed scene, said program comprising instructions for implementing the steps of a method according to the first aspect of the invention during execution of the program by module processing means According to a third aspect, the present invention proposes a module for analyzing a dynamic scene observed using a sensor block (s) comprising a step of defining a grid cut into cells and corresponding to the observed scene, said device being adapted for, s at the base of a grid cut into cells and corresponding to the observed scene, iteratively implement the processing of: - collecting at least one new observation of the sensor block (s) at a kth iteration and determining, as a function of the new observation collected, a first probability of occupation of each cell of the grid and modeling the operation of the sensor block (s); determining, for each cell, at the kth iteration, a second probability of occupation of the cell and a set of probability (s) of movement of the content of the cell as a function of the first probability of occupation of the determined cell at the kth iteration, said analysis module being characterized in that it is adapted to determine said second probability of occupation of the cell and the set of probability (s) of movement of the contents of the cell according to the first probability of occupation of the determined cell at the kth iteration by solving equation 13 (07ZG) = Li10 0-1VV-1P (CA 0-11717-1Z) where C: identifier of the cell in question; A: identifier of the cell which contained, at the (k-1) th iteration what contains the cell considered at the kth iteration; O: state of occupation of the cell considered, among the states "empty" and "busy"; 0-1: state of occupation of cell A at the (k-1) th iteration; V: speed of the cell C considered; V-1: speed of cell A at the (k-1) th iteration; Z: observations of the sensors from the first iteration to the kth iteration; said module being adapted to determine, at the kth iteration, the respective speed and position of a set of fictitious particles in the grid as a function of the velocities, of the positions of the particles determined at the (k-1) th iteration and the probability P (V1V1); said module being adapted to determine particles localized in each cell as a function of the determined positions and to split the resolution of the equation, for a UAE-1V-1 P (C'il0 0 -1171 / 7-Z) cell, into the resolution of a static part corresponding to P (O = empty, V = 01ZC) and P (O = occupied, V = 01ZC) and in the resolution of a dynamic part corresponding to P (O = occ, V = i = 1 to nk, where nk is the number of particles determined in cell C for the kth iteration, said module being adapted to determine the static part of cell C at the kth iteration as a function of the static part of cell C determined at the (k-1) th iteration and P (010-1) and to determine the probability of P (O = occupied, V = vk, 1ZC) from the dynamic part of cell C to the kth iteration based of the probability P (O = occupied, V = -1) -1,1ZA) calculated at the (k-1) th iteration for the dynamic part of cell A and P (010-1), where a particle pi determined in cell C at the kth iteration with a velocity was in cell A at the (k-1) th iteration with a velocity -1) -1.
    [0003] These features and advantages of the invention will become apparent on reading the description which follows, given solely by way of example, and with reference to the appended drawings, in which: FIG. 1 represents an occupancy grid and FIG. transition histogram of a grid cell; Figure 2 shows a view of a vehicle and a grid covering a scene around the vehicle in one embodiment of the invention; FIG. 3 represents an analysis device in one embodiment of the invention; FIG. 4 represents a set of steps implemented in one embodiment of the invention; FIG. 5 represents a grid of occupation and the representation of the probabilities of movement associated with a cell of the grid in one embodiment of the invention.
    [0004] According to a usual convention of the domain of probabilities, if U, V, W, any variables, P (U) designates the probability of obtaining U, and P (UIV) indicates the probability of obtaining U, the value of V being known. Moreover, the expression of the joint probability, which allows the estimation of the probability functions of combinations of variables, is the following: P (UVW) P (W1 U) w P (UVW) A BOF algorithm estimates, during successive iterations, and using a spatial grid decomposed into cells and representing the analyzed scene, the occupation of the cells of the grid and the movement of the contents of the cells, using probabilities function of random variables representing the magnitudes to be studied, which are associated with each cell and whose evolutions over the iterations describe the unfolding of the observed scene (see in particular Pierre Bessière, Emmanuel Mazer, Manuel Juan Ahuactzin-Larios, and Kamel Mekhnacha, "Bayesian Programming ", CRC Press, December 2013, or MK Tay, Kamel Mekhnacha, Chen Cheng, and Manuel Yguel," International Effective Day of the Bayesian Occupational Filter for Target Tracking in Dynamic Environments, "International Day nal of Vehicle Autonomous Systems), 6 (1): 155 {171, 2008). The variables considered in the present case are described below: C: identifier of the cell in question; - A: identifier of the "antecedent": the antecedent is the cell which contained, at the previous iteration, what contains the cell considered at the current iteration; 0: occupation of the considered cell; occupation takes a state among "empty" and "busy"; 0-1: occupation of the antecedent at the previous iteration; V: velocity (dimension vector 2) of the cell in question; V-1: speed of the antecedent to the previous iteration Z: observations of the sensors since the first iteration until the current iteration. The joint probability is expressed using equation (1) below: P (CA 0 0-1 V V-1 Z) = P (A) P (0-1V-11A) P (0) V10-1 V-1) P (C 1A V) P (ZOVC), where: the probabilities P (A) represent the distribution on all possible antecedents A; it is, in the embodiment considered, chosen uniform, a cell being considered as reachable from all the other cells of the grid with an equal probability; F, (01 v ", 1 A) is the conditional joint distribution on the occupation and the movement of the antecedents, the probabilities corresponding to this distribution are computed at the iteration preceding the iteration considered, P (0 V 1 0- 1 V-1) is the dynamic model of prediction, the chosen model can be decomposed as the product of two terms: P (01 0-1) P (V1 V-10), where: P (01 0-1) is the conditional distribution on the occupation of the cell considered, the occupation of the antecedent being known, this 1-ee distribution is defined by the transition matrix T, T = e 1-e which authorizes the change of state of occupancy of a cell with a low probability to take into account evaluation and observation errors P (V 1 V-10) is the conditional distribution on the speed of the current iteration, the speed at the previous iteration being known, usually it follows a normal law, centered on the speed at the iteration echoes to represent a model of Gaussian acceleration; this distribution can be adapted to correspond to different objects observed; and since it is considered that an empty cell can not move, a Dirac function is added to prohibit movement in this case; - P (C 1 A V) is the distribution which indicates whether the cell C considered is accessible from the known antecedent A = a and with the known velocity V = v; this distribution is a function of Dirac with a value equal to 1 if and only if a + v.dt C C, where dt is the step of time separating the current iteration of the previous iteration; - P (Z 1 O V C) is the distribution on the observation values of the sensor, also called "sensor model". An algorithm BOF estimates, during successive iterations, the occupation of the cells and the speed of the contents of the cells according to current observations provided by the sensor (s) for each cell, by the determination of P ( OV1ZC).
    [0005] By discretizing the antecedent cells and velocities, P (OV1ZC) can be written using the following equation (2): P (OVIZC = ZAD-1L-1P (CAO 0-1171 / -1Z) ) LAO G-117V-P (CA 0 -117V-1Z) In Figure 2, a vehicle 10 is shown in plan view. The vehicle 10, in addition to the usual means of locomotion specific to vehicles, comprises a sensor unit 11 and an analysis device 20 in one embodiment of the invention.
    [0006] As shown diagrammatically in FIG. 3, the analysis device 20, also called HSBOF module (for "Hybrid Sampling BOF") comprises a microprocessor 21 and a memory 22. The memory 22 comprises software instructions, which when executed by the microprocessor 21 implements the steps indicated below and performed by the analysis device 20. In the case considered, the analysis device 20 further comprises a parallel compute accelerator adapted to process in parallel operations implemented implemented in the method detailed below. The sensor unit 11 is adapted on the one hand to estimate the movements of the vehicle 10, in the form of translations and rotations. These displacement estimates are provided to the analysis device 20. The sensor unit 11 further comprises an on-board sensor (or several on-board sensors) making it possible to take measurements and, depending on these measurements, to detect the presence of occupied zones. for example, assimilated according to certain criteria to obstacles in the environment of the vehicle 10 and their positions relative to a frame of reference (X1, X2) related to the vehicle (it will be noted that in other embodiments, a depth map is determined by these measures). Such a sensor comprises for example a laser rangefinder with a source emitting a laser beam and which scans a surface cp in the plane (X1, X2), and / or a stereo camera, etc. and detects the presence of obstacle (s) in the scanned scene, and their position according to the analysis of reflected beams received by the range finder. This scene observation data comprising the position of detected obstacles are provided by the sensor 11 to the analysis device 20.
    [0007] In embodiments, scene observation data indicating the presence of obstacle (s) in the scanned scene, and their position, are provided to the analysis device 20 by one or more external sensors 13 to the vehicle 10 for example a sensor attached to the edge of the road on which the vehicle 10 is traveling. A space grid 14 corresponding to a portion of the environment of the vehicle 10 is also considered, for example a view from above of a zone of width and length fixed in the plane (X1, X2), for example of 20 meters each. , which extends for example from the vehicle forward towards the longitudinal axis X1 of the vehicle. This gate 14 is cut into C cells, for example 100,000 cells, so as to form a spatial grid representing a top view of the vehicle environment.
    [0008] The analysis device 20 is adapted to deduce, at each instant tk = to + k.dt, these observation data of the vehicle environment received from the sensor (s) 11, the distribution on the observation values received , also called "probabilistic sensor model": P (ZOVC). So either zk the observations of the environment of the sensor block 11 to consider for the moment tk. The probability P (ZOVC) determined at the instant tk for each cell C of the grid 14 by the analysis module 20 as a function of the observations z1 to zk received at the iteration instants t1 to tk models the behavior of the sensor 11 and takes into account the sound of the measurement by the sensor block 11, the imprecision factors such as temperature, lighting, the colors of the objects, the masking of objects, the ability of the sensor to detect or not the objects and to detect objects when there are none ("false detections"). It will be noted that the observation Z does not take in the case considered the shape of the coordinates of the objects detected in the reference linked to the vehicle and axes X1, X2, but takes the value 1 in a cell C where an obstacle has been detected. and 0 elsewhere. The probabilistic sensor model is, depending on the case, estimated from a modeling of the physical characteristics of the sensor and the objects to be detected, or "learned" experimentally. Examples of sensor models are for example described in FR 0552736 or at http: //ieeexplore.ieee.orq/stamp/stamp.isp Tp = & arnumber = 30720 & tab = 1.
    [0009] In this case, considering that the analysis module 20 collects at an instant tk updated observation data provided by m sensors, each providing observations, j = 1 to m, the observation vector zk = (z1, z2 , zm). Typically, a laser gives an observation corresponding to a list of impact distance values for different directions. The analysis module 20 obtains at each instant tk an updated observation vector zk. From these instantaneous vectors, the analysis device 20 deduces from zk and from the "sensor model" (in this case here multisensor) the probability P (Z = zk OVC), which gives an instantaneous grid (indicating for each C, if the box were occupied, and with such speed, what would be the probability of having these observations) which is then provided as input data of step 103 detailed below and in which the probabilities P (OV z '... zk C) are calculated recursively using the estimates P (O VI z' ... zk- 'C) calculated at tk-i.
    [0010] The invention is based on the following principles: decomposition of each cell into a static part corresponding to the unoccupied part of the cell and to the part of the cell occupied by objects of zero velocity, and in a dynamic part corresponding to the part the cell occupied by moving objects; represent the motion distribution of a cell of the occupancy grid by a set of particles representing the dynamic part of the cell, instead of a regular motion distribution grid, which allows a more compact representation. Thus the determination of P (OV1ZC) amounts to calculating the probability associated with each particle P (O = occ, V = vk I ZC) and the probabilities associated with the static domain P (O = occ, V = 0 I ZC) and P (O = empty, V = 0 ZC). Using decomposition (1) and considering the denominator of equation (2) as a scale factor, this equation (2) can be rewritten as follows: P (OVIZC) "P (ZIOC) P ( A) P (0 -V1A) P (00-1) P (VOV-1) P (CIAV) where the symbol "" "means" is proportional to ". (3) We denote a (o, v) = P (A) P (0-117-1k1) P (o10-1) P (+ 17-1) P (CIAv) (4) A0-117-1 where oa for value the occupation status "busy" (named "occ") or "empty", and v is a speed. Method 100 A set 100 of steps 101-106 implemented in a kth iteration at time tk = to + k.dt, by the analysis device 20 is described below with reference to FIG. embodiment of the invention (kk1). It is considered that at the end of the previous iteration, ie (k-1) th iteration, for each cell C of the grid, the probabilities associated with the static domain were obtained: P (O = occ, V = 0 ZC), named Pk-1 (0 = occ, V = 0 ZC), P (O = empty, V = 0 I ZC), named Pk-1 (0 = empty, V = 0 ZC), and a set of particles has been assigned to cell C, such that each particle in this set has a respective speed -1) -1 and a respective position .x-1 and the probabilities associated with the dynamic domain have been obtained. P (0 = occ, V = k-1 II ZC), named pk1 (0 = OCC, V = 1) -1 ZC). In the embodiment considered, the position .x-1 makes it possible to position the particles exactly within the cell. The introduction of precise particle positions within a cell makes it possible to solve spatiotemporal aliasing problems. In a step 101 for determining the particle propagation, for each particle, the motion model is applied, according to the distribution P (VI V-1): k 'k-1 V, = V, + G k' k -1 k '= di. V, where o is a Gaussian noise centered at 0 of covariance E. In a step 102, the displacement of the gate corresponding to a displacement of the vehicle 10, if any, is taken into account, as a function of the rotation Rk and the translation Tk performed between the iteration (k-1) and the current iteration k and determined by the sensor block 11 :: kk 'V, = Rk. xk Rk xk '+ Tk In a step 103, the observations zk of the scene to be considered for the moment tk are collected (sub-step 103_0) and the probabilities P (ZIO, V, C) resulting from these observations zk and modeling the sensor, as described above, are deduced from these observations.
    [0011] Then the occupancy and motion probabilities taking into account these observations collected are updated (substeps 103_1 to 103_3 detailed below). The static part of each cell C is updated, in the sub-step 103 1, using the static part of the considered cell C determined at the end of the (k-1) th iteration, ie Pk-1 ( occ, 0 Z, C) and Pk-1 (empty, 01Z, C). Indeed, P (0-1 = occ, 17-1 = 01Z, A = C) = Pk-1 (occ, 0 IZ, C), and P (0-1 = empty, V-1 = 0 Z, A = C) = Pk-1 (empty, 0 Z, C). It follows that the coefficients a (occ, 0) and a (empty, 0) according to the definition given in equation (4) are obtained for the cell C considered by the following calculation: a (occ, 0) = Pk-1 (occ, 0 IZ, C). (1-e) + Pk-1 (empty, 0 IZ, C) .ca (empty, 0) = Pk-1 (occ, O, Z, C). c + Pk-1 (empty, 0 1Z, C). (1-c), where c = P (O = occ 0-1 = empty) = P (0 = empty0-1 = occ). The dynamic part of the cells is in turn updated, considering the particles. Thus, for each cell C of the gate, in a step 103_2: - it is determined which particles, among the sets of particles associated with the cells, are found in the cell C after the implementation of steps 101 and 102. Consider that nck particles are found in the considered C cell. Each of these particles has a respective velocity vk and a current position xk within the cell C and was in a cell, named a, at the iteration k-1, as determined at the end of step 102; and we define the associated weighting coefficient 142-1 for this particle: 142-1 = P (0-1 = occ, V-1 = vk 1 1 A = a ,, z1, ..., zk-1) which is equal to Pk-1 (0 = occ, V = -1) -1 IZ, a,), determined at the iteration k-1 (substep 103_3) relative to the cell a, - the coefficient a ( occ,) according to the definition given in equation (4) is then determined for cell C considered as equal to: a (occ,) = w, k-1. (1 C) ; an empty cell is assumed to have zero velocity: a (empty, vik) = o. In a substep 103_3 for updating the probabilities, for each cell C, to complete the resolution of equation (3), we calculate [3 (occ, 0), [3 (empty, 0) and [3 (occ, vk) in the following manner, using the updated observation probabilities P (ZIO, V, C) resulting from the substep 103_0 and the coefficients a (occ, 0), a (empty, 0 ) and a (occ,))) calculated during substeps 103_1 and 103_2: [3 (occ, 0) = P (ZI occ, 0, C) .a (occ, 0) [3 (empty, 0) = P (empty ZI, 0, C) .a (empty, 0) [3 (occ, vik) = P (ZI occ, vik, C) .a (occ, vik).
    [0012] The denominator of equation (2), named scaling factor "d", is calculated: d = [3 (occ, 0) + [3 (empty, 0) + E, 3 (occ, vk), where E, [3 (occ,) is the sum of the coefficients [3 (occ,) calculated for the particles in cell C.
    [0013] Finally, we obtain for cell C the following probabilities which represent I "Bayesian filtering occupation and the associated probabilities of movement: c Pk (occ, 0 IZ, C) - fl (odc, 0) for zero velocity, Pk ( occ,) 6 (occ, v1 :) IZ, C) - for each speed vik, the velocities vk being the velocities of the particles determined in the cell C, Pk (empty, 0 IZ, C) - 13 (empty, 0 ) A step 104 (optional) of reorganization of the particles is then implemented to adapt the number of particles per cell and for example to duplicate the high probability particles, in order to allow a better efficiency of the set of steps 100, implementing the Proc_reorg process indicated below It will be noted that the accuracy of the representation of the velocity distribution in a cell increases with the number of particles in the cell. , the total number of particles, named nbPart , is kept constant over the iterations (the computation load is thus the same for each iteration, the memory to be allocated is constant and known). This total number may change in other embodiments (for example, the total number could be increased for a very dynamic scene and decreased for a more static scene).
    [0014] We consider for each cell C the weighted set of nck particles defined by the nck triplets {xk, vk, wl each associated with a particle, where wk = Pk (O = occ, V = vk Z, C). According to the process Proc reorg, a number equal to nbPart is repeated the set of successive successive steps i, ii and iii, after having initialized to an empty set, per cell of the grid, a new set of particles attributed to the cell. : i / draw a cell identifier according to a law of occurrence proportional to the weight of the cell (the weight of a cell C is equal to the sum of the nck weight wk of the particles in the cell and therefore represents Pk (0 = occ , V # 01Z, C)); ii / randomly selecting a particle from the nck particles of this cell, according to a law of occurrence proportional to the weight wk of the particle; iii / add the particle selected in ii / to the new set of particles allocated to the cell whose identifier has been drawn in i /. Once these steps are repeated, the weight of the particles is normalized per cell, i.e. the weight of each particle in the new set of particles assigned to the cell is divided by the sum of the weights of the particles in this new set; it is these normalized weights that are considered as wk weights for the next iteration. In a step 106, the value of k is increased by 1.
    [0015] When a new object appears on the grid, it is necessary to initialize one or more corresponding particles with a speed distribution. For this, in one embodiment, a constant occurrence probability with a uniform distribution between busy / empty and static / dynamic states is added. For example, in step 103 1, we calculate a (occ, 0), respectively a (empty, 0), by adding pa / 4, respectively pa / 2, or a (occ, 0) = Pk-1 ( ## STR2 ## (1-c) + Pk-1 (vacuum, O, C) .0 + Pa / 4a (void, O) = Pk-1 (OCC, O, Z, C). + Pk-1 (empty, 0 Z, C). (1-c) + pa / 2; and in step 103 2, the cell is considered to have an additional dynamic term a (occ, V = "unknown") = pa / 4. Thus, in step 104 of reorganizing the particles, when the "unknown" speed is selected (step ii /), the speed is determined from a uniform distribution, uniformly selecting a speed on [-V ', + V 'x] x [-V'x, + V'x] and associating it with the cell. In an embodiment that may or may not combine with the previous one, to prevent particles from tracking a static object, each particle contributes to the static portion represented by a (occ, 0) calculated in step 103 1 by a coefficient of value decreasing with the speed standard: a (occ, 0) = Pk-1 (occ, 01Z, C). (1-c) + 13k-1 (empty, 01Z, C) .c + at e 2cr >. 142-1 * (1- e); -0v = `and in step 103_2: a (occ,) = (1- e 2 ° 1). "142-1. (1 - £), where as is a constant coefficent representing the static velocity contribution (for 111) k> 3 <as, the static contribution is insignificant).
    [0016] In an initialization step prior to the first iteration (k = 1) of the set of step 100, it is proceeded, in one embodiment, as for the "unknown" type of antecedents: uniform draw of the speed on [- Vmax, Vmax] x [- Vmax, Vmax]. The position is taken by pulling the grid evenly (drawing a floating position on [- gridWidth / 2, gridWidth / 2] x [-gridHeight / 2, gridHeight / 2], where gridWidth is the grid width and gridHeight is the height of the grid), the occupation is fixed at 0.5 (unknown). The w ° are equal to 1 / nc °, this number ng of particles in cell C being a result of the previous random draw. The initial static probability is a parameter with little impact because the re-evaluation takes place very quickly.
    [0017] Thus according to the invention, the conventional representation BOF comprising a occupation grid and associating with each cell of the grid a speed histogram is replaced by a occupation grid with, for each cell, a probability of occupation, a coefficient statistic corresponding to a probability of zero velocity, a set of particles in the cell, each having a respective velocity v, and a set of probabilities P (V = y). The invention makes it possible to construct at each iteration, via the grid, a discrete probabilistic map of the scene, distinguishing the static and dynamic components. The motion of dynamic components is estimated using particles. The quality of the occupancy and speed estimates is improved, by better precision of the calculated speeds and a decrease of the spatio-temporal aliasing effects. By way of illustration, the occupancy gate 14 obtained in one embodiment of the invention is shown. The cell 15 of the occupancy gate 14 comprises for example 3 pu particles, at respective positions in the cell, having respective speeds seen, j = 1 to 3.
    [0018] The observations and estimates of displacement of the vehicle are made by the sensors and supplied to the analysis device 20, then processed in real time by the analysis device 20. For example, the iteration step dt is equal to 0.06 second ., the grid refreshing observations and displacement estimates being provided to the analysis device 20 in time steps of time less than or equal to dt. The occupancy and velocity values associated with the particles determined for each cell at each iteration may, in embodiments of the invention, be exploited in a different manner in a step 105: current action / decision as a function of the values of occupations and velocity fields determined for each cell (by prediction of future state (s) of the environment (for example risk assessment of collisions between the vehicle 10 and objects corresponding to cells whose probability of occupancy is greater than a given occupation threshold and the observed distance is below a given distance threshold), then for example current action / decision according to one or more determined future states (automatic braking of the vehicle For example), it will be noted that the velocity distributions inferred by the invention are directly usable for predicting the evolution of the scene, and therefore to deduce the int ersection between the vehicle and obstacles, including dynamic objects. The steps of the method 100 are easily parallelizable. The determination of the occupancy grid and the speed fields is feasible for an analyzed environment comprising all sorts of objects previously known or not known using sensors that are even heterogeneous, thanks to the use of probabilistic sensor models. The present invention makes it possible to significantly reduce the computing and memory resources required for the determination of cell occupation and motion probability distributions. It makes it possible to achieve gain by a factor of 100. It makes it possible to compensate the own motion of the sensors (to calculate the rotation / translation operations of the grids and the particles on the velocity fields when they were considered in the form of histograms BOF was in practice impossible due to excessive aliasing and 4D interpolation too expensive in calculations).
    权利要求:
    Claims (11)
    [0001]
    CLAIMS1.- A method of analyzing a dynamic scene observed using a block (11) of sensor (s) comprising a step of defining a grid (14) cut into cells (C) and corresponding to the observed scene; and iteratively computer-implemented steps of: collecting at least one new observation of the sensor block (s) at the kth iteration and determining, based on the new observation collected, a first probability of occupation of each cell of the grid modeling the operation of the sensor block (s); determining, for each cell, at the kth iteration, a second probability of occupation of the cell and a set of probability (s) of movement of the content of the cell as a function of the first probability of occupation of the cell determined at the keme iteration, said method being characterized in that the second probability of occupation of the cell and of the set of probability (s) of movement of the content of the cell as a function of the first probability of occupation of the cell determined at the kth iteration is determined by solving the equation p (o 1v-1P (CAO 0-1171 / 7-1-Z) - v izc) = A ° the speed and the respective position of a set of fictitious particles in the grid 30 are determined at the kth iteration as a function of the velocities, positions of the particles determined at the (k-1) th iteration and the probability P (V1V-1); the method comprises a step of determining the localized particles in each cell according to the determined positions and in that the resolution of the equation, for a cell, is split into the resolution of a static part corresponding to 35 P (0 = empty, V = 01ZC) and P (0 = occupied, V = 01ZC) and in the resolution of a portion AO 0 - ivv-i P (CAO 0 -1171 / 7-1Z) where C: identifier of the considered cell ; A: identifier of the cell which contained, at the (k-1) th iteration what contains the cell considered at the kth iteration; O: state of occupation of the cell considered, among the states "empty" and "busy"; 0-1: state of occupation of cell A at the (k-1) th iteration; V: speed of the cell C considered; V-1: speed of cell A at the (k-1) th iteration; Z: observations of the sensors from the first iteration to the kth iteration, corresponding to the dynamics of P (O = occ, V = i = 1 to nk, where nk is the number of particles determined in cell C for the kth iteration; the static part of the cell C at the kth iteration being determined as a function of the static part of the cell C determined at the (k-1) th iteration and P (010-1), the probability of P (O = occupied , V = vk, 1ZC) of the dynamic part of the cell C being determined at the kth iteration as a function of the probability P (O = occupied, V = -v-1,1ZA) calculated at the (k-1) th iteration for the dynamic part of cell A and P (010-1), where the particle p, determined in cell C at the kth iteration with a velocity was in cell A at the (k-1) th iteration with a speed
    [0002]
    2. A method according to claim 1, wherein at the end of the kth iteration, one or couples (p ,,)), where xk is the position of the particle p, at the kth iteration, are duplicated within of a cell (C) or deleted so that the number of pairs per cell is a function of the dynamic part determined for the cell.
    [0003]
    3. A method according to claim 2, including a step of selecting the pair to be duplicated / deleted, the selection of said pair being performed as a function of the probability P (O = occupied, V = -vIZC) determined at the kth iteration, where vk is the speed component of the couple.
    [0004]
    4. A method according to claim 2 or 3, wherein the total number of particles in the grid (14) is constant over the iterations.
    [0005]
    5.- Method according to one of the preceding claims, wherein P (O = empty, V = 01ZC) determined at the kth iteration and noted Pk (0 = empty, V = 01ZC), is determined according to the product of a first term which is a function of the first probability of occupation and a second term, said second term being a function of: - Pk-1 (0 = occ, V = 01ZC). (1-e) + Pk-1 (0 = empty, V = 01ZC) .c, and / or - coeff (). Pk-1 (0 = busy, V = 1), 1ZA). (1 - where coeff (v is a decreasing function of 11-v11, and / or - a probability of occurrence pa of a new object in the scene observed.
    [0006]
    6. Computer program to be installed in an analysis module (20) of a dynamic scene observed using a sensor block (11) comprising a step of defining a grid ( 14) cut into cells (C) and corresponding to the scene observed, said program comprising instructions for implementing the steps of a method according to one of claims 1 to 5 during execution of the program by means of processing of the analysis module.
    [0007]
    7. Analysis module (20) of a dynamic scene observed using a block (11) of sensor (s) comprising a step of defining a grid (14) cut into cells (C) and corresponding to the observed scene; said device being adapted to, on the basis of a grid (14) cut into cells (C) and corresponding to the observed scene, iteratively implement the processing of: - collecting at least one new observation of the sensor block (s) ) a kth iteration and determine, based on the new observation collected, a first probability of occupation of each cell of the grid and modeling the operation of the sensor block (s); determining, for each cell, at the kth iteration, a second probability of occupation of the cell and a set of probability (s) of movement of the content of the cell as a function of the first probability of occupation of the determined cell at the kth iteration, said analysis module being characterized in that it is adapted to determine said second probability of occupation of the cell and the set of probability (s) of movement of the contents of the cell according to the first probability of occupation of the determined cell at the kth iteration by solving the equation P- (Q EA0-1v- P (CAC 0-11717 Z) I / IZ C) = v, LAO 0-117V-1 P ( CA00-11717-12) where C: identifier of the considered cell; A: identifier of the cell which contained, at the (k-1) th iteration what contains the cell considered at the kth iteration; O: state of occupation of the cell considered, among the states "empty" and "busy"; 0-1: state of occupation of cell A at the (k-1) th iteration; V: speed of the cell C considered; V-1: speed of cell A at the (k-1) th iteration; Z: observations of the sensors from the first iteration to the kth iteration, said module being adapted to determine, at the kth iteration, the speed and the respective position of a set of fictitious particles in the grid as a function of the speeds, positions of the particles determined at the (k-1) th iteration and the probability P (VIV-1) 'said module being adapted to determine localized particles in each cell according to the determined positions and to split the resolution of the equation , for a cell, in the resolution of a static part corresponding to P (0 = empty, V = ORC) and P (O = occupied, V = OIZC) and in the resolution of a dynamic part corresponding to P (0) = occ, V = v,, IZC), i = 1 to nk, where nk is the number of particles determined in cell C for the kth iteration; said module being adapted to determine the static part of the cell C at the kth iteration as a function of the static part of the cell C determined at the (k-1) th iteration and P (0104); and to determine the probability of P (O = occupied, V = VIk, IZC) from the dynamic part of cell C to the kth iteration as a function of the probability P (O = occupied, V =, IZA) calculated at k-Dth iteration for the dynamic part of cell A and P (010-1), where the pi particle determined in cell C at the kth iteration with viic velocity was in cell A at (k-1 ) iteration with a speed.
    [0008]
    8.- Analysis module according to claim 7, adapted for, at the end of the kth iteration, duplicate or delete one or couples (pi, (xik)), where xi: is the position of the pi particle to the kth iteration, within a cell (C) so that the number of pairs per cell is a function of the dynamic part determined for the cell.
    [0009]
    9. Analysis module according to claim 8, adapted to select said pair to duplicate / delete, as a function of the probability P (0 = occupied, V = vik IZC) determined at the kth iteration, where 1, /, is the speed component of the couple.
    [0010]
    10.- Analysis module according to claim 8 or 9, wherein the total number of particles in the grid (14) is constant over the iterations.
    [0011]
    11.- Analysis module according to one of claims 7 to 10, adapted to determine P (O = empty, V = 01ZO) at the kth iteration, noted Pk (0 = empty, V = 01ZO), depending on the produces a first term which is a function of the first probability of occupation and a second term, said second term being a function of: - Pk-1 (0 = occ, V = 01ZO). (1-e) + Pk- 1 (0 = empty, V = 01ZO) .c; and / or - coeff (vk). Pk-1 (0 = occupied, V = 1), 1ZA). (1 - where coeff (vk) is a decreasing function of Ovk, and / or - a probability of occurrence pa of a new object in the observed scene 15
    类似技术:
    公开号 | 公开日 | 专利标题
    FR3022049A1|2015-12-11|METHOD FOR ANALYZING A DYNAMIC SCENE, ANALYSIS MODULE AND COMPUTER PROGRAM THEREOF
    Rummelhard et al.2015|Conditional monte carlo dense occupancy tracker
    CN101283387A|2008-10-08|Vehicle steering aid method and improved related device
    FR3020616A1|2015-11-06|DEVICE FOR SIGNALING OBJECTS TO A NAVIGATION MODULE OF A VEHICLE EQUIPPED WITH SAID DEVICE
    EP2537146A1|2012-12-26|Video surveillance system
    EP3252615B1|2021-09-08|Method and system for determining cells crossed by a measuring or viewing axis
    EP2297564A1|2011-03-23|Method and device for frequency analysis of data
    EP3364213A1|2018-08-22|Method and system for contextualised perception of material bodies
    Sless et al.2019|Road scene understanding by occupancy grid learning from sparse radar clusters using semantic segmentation
    Benhaddou et al.2014|Multichannel deconvolution with long-range dependence: A minimax study
    EP3189389B1|2021-09-08|Location and mapping device with associated method
    CN112906816B|2021-11-09|Target detection method and device based on optical differential and two-channel neural network
    Halimi et al.2021|Robust and guided bayesian reconstruction of single-photon 3d lidar data: Application to multispectral and underwater imaging
    EP3745158B1|2022-02-16|Methods and systems for computer-based determining of presence of dynamic objects
    US20200033870A1|2020-01-30|Fault Tolerant State Estimation
    EP3384462B1|2020-05-06|Method for characterising a scene by calculating the 3d orientation
    Kullberg et al.2021|Online Joint State Inference and Learning of Partially Unknown State-Space Models
    FR2977679A1|2013-01-11|METHOD AND DEVICE FOR DETECTING A TARGET MASKED BY HIGH ENERGY REFLECTORS
    CN112508803A|2021-03-16|Denoising method and device for three-dimensional point cloud data and storage medium
    WO2020202505A1|2020-10-08|Image processing apparatus, image processing method and non-transitoty computer readable medium
    De Maio et al.2022|Deep Bayesian ICP Covariance Estimation
    Naghiu et al.2009|Influence of driver behavior patterns in correcting video sensing errors in traffic surveillance applications
    WO2021162018A1|2021-08-19|System and method for tracking expanded state of moving object with model geometry learning
    FR3100641A1|2021-03-12|DETERMINATION OF REAL-TIME ENVIRONMENTAL INFORMATION AND SELECTIVE REGRESSION, FOR A SYSTEM
    CN111627056A|2020-09-04|Depth estimation-based driving visibility determination method and device
    同族专利:
    公开号 | 公开日
    CA2950847A1|2015-12-10|
    US20170256071A1|2017-09-07|
    JP6550403B2|2019-07-24|
    WO2015185846A1|2015-12-10|
    JP2017525015A|2017-08-31|
    KR20170055953A|2017-05-22|
    EP3152675A1|2017-04-12|
    FR3022049B1|2016-07-22|
    US10290116B2|2019-05-14|
    引用文献:
    公开号 | 申请日 | 公开日 | 申请人 | 专利标题
    WO2007028932A1|2005-09-09|2007-03-15|Institut National De Recherche En Informatique Et En Automatique|Vehicle steering aid method and improved related device|
    FR2890773A1|2005-09-09|2007-03-16|Inst Nat Rech Inf Automat|Driving assistance method, e.g. for ABS and ACC systems, in which an area in front of a vehicle is discretized as a multiplicity of cells and the vehicle collision probability is determined based on individual cell calculations|
    EP2289754A1|2009-08-31|2011-03-02|Toyota Motor Europe NV/SA|Vehicle or traffic control method and system|
    FR552736A|1922-06-10|1923-05-05|Removable spark plug|
    DE19641261C1|1996-10-07|1998-02-12|Siemens Ag|Distance sensor measuring quality evaluation method|
    US8538675B2|2009-06-01|2013-09-17|Raytheon Company|Non-kinematic behavioral mapping|
    JP5206752B2|2010-08-30|2013-06-12|株式会社デンソー|Driving environment recognition device|
    DE102013223803A1|2013-11-21|2015-05-21|Robert Bosch Gmbh|Method and device for segmenting an occupancy grid for an environment model of a driver assistance system for a vehicle|
    DE102014208009A1|2014-04-29|2015-10-29|Bayerische Motoren Werke Aktiengesellschaft|Capture static and dynamic objects|
    FR3022049B1|2014-06-06|2016-07-22|Inria Inst Nat De Rech En Informatique Et En Automatique|METHOD FOR ANALYZING A DYNAMIC SCENE, ANALYSIS MODULE AND COMPUTER PROGRAM THEREOF|FR3022049B1|2014-06-06|2016-07-22|Inria Inst Nat De Rech En Informatique Et En Automatique|METHOD FOR ANALYZING A DYNAMIC SCENE, ANALYSIS MODULE AND COMPUTER PROGRAM THEREOF|
    JP2017068817A|2015-10-01|2017-04-06|三星電子株式会社Ｓａｍｓｕｎｇ Ｅｌｅｃｔｒｏｎｉｃｓ Ｃｏ．，Ｌｔｄ．|Method and apparatus for modeling object|
    US11132611B2|2016-05-27|2021-09-28|Kabushiki Kaisha Toshiba|Information processing apparatus and information processing method for determining presence probability of object|
    US11222438B2|2016-05-27|2022-01-11|Kabushiki Kaisha Toshiba|Information processing apparatus, vehicle, and information processing method for presence probability of object|
    US10195992B2|2017-04-03|2019-02-05|Ford Global Technologies, Llc|Obstacle detection systems and methods|
    DE102017206436A1|2017-04-13|2018-10-18|Bayerische Motoren Werke Aktiengesellschaft|Method for detecting and / or tracking objects|
    US11255959B2|2017-06-02|2022-02-22|Sony Corporation|Apparatus, method and computer program for computer vision|
    DE102017217156B4|2017-09-27|2021-12-23|Robert Bosch Gmbh|Method and device for controlling a driver assistance system using a stereo camera system with a first and a second camera|
    US11048927B2|2017-10-24|2021-06-29|Waymo Llc|Pedestrian behavior predictions for autonomous vehicles|
    US11077756B2|2017-11-23|2021-08-03|Intel Corporation|Area occupancy determining device|
    FR3087559B1|2018-10-17|2020-11-20|Transdev Group|ELECTRONIC DEVICE AND METHOD FOR MONITORING A SCENE VIA AN ASSEMBLY OF ELECTRONIC SENSOR , ASSOCIATED COMPUTER PROGRAM|
    FR3089671B1|2018-12-11|2020-11-06|Psa Automobiles Sa|Method for determining a current value of an occupancy parameter relating to a portion of a space located near a land motor vehicle|
    DK180694B1|2019-01-08|2021-12-02|Motional Ad Llc|FIELD THEORY-BASED PERCEPTION FOR AUTONOMIC VEHICLES|
    US20200257931A1|2019-01-08|2020-08-13|Aptiv Technologies Limited|Field theory based perception for autonomous vehicles|
    WO2021049231A1|2019-09-11|2021-03-18|日立オートモティブシステムズ株式会社|Occupied grid map management device|
    EP3901661A1|2020-04-22|2021-10-27|Bayerische Motoren Werke Aktiengesellschaft|Method and control unit for operating an autonomous vehicle|
    法律状态:
    2015-06-04| PLFP| Fee payment|Year of fee payment: 2 |
    2015-12-11| PLSC| Publication of the preliminary search report|Effective date: 20151211 |
    2016-06-29| PLFP| Fee payment|Year of fee payment: 3 |
    2017-06-19| PLFP| Fee payment|Year of fee payment: 4 |
    2018-06-25| PLFP| Fee payment|Year of fee payment: 5 |
    2020-06-25| PLFP| Fee payment|Year of fee payment: 7 |
    2021-06-29| PLFP| Fee payment|Year of fee payment: 8 |
    优先权:
    申请号 | 申请日 | 专利标题
    FR1455183A|FR3022049B1|2014-06-06|2014-06-06|METHOD FOR ANALYZING A DYNAMIC SCENE, ANALYSIS MODULE AND COMPUTER PROGRAM THEREOF|FR1455183A| FR3022049B1|2014-06-06|2014-06-06|METHOD FOR ANALYZING A DYNAMIC SCENE, ANALYSIS MODULE AND COMPUTER PROGRAM THEREOF|
    KR1020177000190A| KR20170055953A|2014-06-06|2015-06-02|Dynamic scene analysis method, and associated analysis module and computer programme|
    US15/316,779| US10290116B2|2014-06-06|2015-06-02|Dynamic scene analysis method, and associated analysis module and computer programme|
    PCT/FR2015/051449| WO2015185846A1|2014-06-06|2015-06-02|Dynamic scene analysis method, and associated analysis module and computer programme|
    EP15732833.7A| EP3152675A1|2014-06-06|2015-06-02|Dynamic scene analysis method, and associated analysis module and computer programme|
    JP2016571280A| JP6550403B2|2014-06-06|2015-06-02|Dynamic scene analysis method, associated analysis module and computer program|
    CA2950847A| CA2950847A1|2014-06-06|2015-06-02|Dynamic scene analysis method, and associated analysis module and computer programme|
    [返回顶部]