Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F30/00—Computer-aided design [CAD]
  • G06F30/20—Design optimisation, verification or simulation
  • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F2111/00—Details relating to CAD techniques
  • G06F2111/10—Numerical modelling

Definitions

• TITLE Process for characterizing the movement of mobile entities and associated computer program product
• the present invention relates to a method for characterizing the movement of mobile entities in an environment.
• the present invention also relates to a computer program product comprising software instructions suitable for implementing such a characterization method.
• the present invention relates to the field of movements of mobile entities in a predetermined environment.
• mobile entities we mean things or beings capable of moving in the environment.
• the mobile entities are pedestrians, cyclists, motorcyclists, automobile drivers, robots, drones, or even animals.
• the term "environment” means a space, for example predefined, in which the mobile entities are able to move.
• the environment is a station hall, an airport, a metro hall, a parking lot, or a forest.
• Determining this set of parameters is particularly complex since these parameters generally vary over time and depending on the situation. Furthermore, if the parameter set is not well defined, the scenario predicted by the simulator is unrealistic. In such a case, the predicted scenario is unusable.
• this method requires knowledge of a prior model capable of providing an acceptable set of parameters.
• this method requires a large number of simulations, by the simulator, to make the evolutionary algorithms converge towards a satisfactory solution making it possible to subsequently generate a good set of calibration parameters.
• the invention proposes a method for characterizing the movement of mobile entities in an environment, which is simpler and faster to use.
• the subject of the invention is a method for characterizing the movement of mobile entities, such as pedestrians, over areas of an environment, the method comprising a learning phase comprising the following steps implemented by a calculator:
• the method further comprising an operating phase comprising the following steps:
• the method comprises one or more of the following characteristics, taken in isolation or in all technically possible combinations: - the method further comprises, during the operating phase, the following steps:
• the method comprises, during the operating phase, a step of predicting the movements of the mobile entities in the environment, from the set of parameter(s) obtained, to obtain a predicted scenario;
• the method further comprises, following the prediction step, the following steps:
• the acceptance stage includes the following sub-stages:
• the training step includes the following sub-steps:
• the method further comprises, between the sub-step of calculating at least one set of averaged parameter(s) and the sub-step of applying, a sub-step of normalizing the average number of entities (s) mobile(s) of each zone to obtain a normalized vector of number of mobile entity(ies) whose algebraic norm is preferably equal to one, during the application step, the algorithm learning is applied to the model from the normalized vector, from the at least one set of averaged parameter(s) and from the adjacency matrix;
• the respective set of parameter(s) obtained during the exploitation phase comprises at least one parameter chosen from a group consisting of: - an average speed of mobile entities in the environment,
• the adjacency matrix comprising a plurality of rows, a plurality of columns, and a coefficient for each row and each column, each coefficient being equal to a predefined value if a displacement is possible from the zone whose number is equal to the column of the coefficient, towards the zone whose number is equal to the line of the coefficient;
• the present invention also relates to a computer program product comprising software instructions which, when executed by a computer, implement such a characterization method.
• FIG. 1 Figure 1 is a schematic view of an environment in which mobile entities are able to move
• FIG. 2 is a schematic view of a computer capable of implementing a characterization method according to the invention.
• FIG. 3 Figure 3 is a flowchart of the characterization method implemented by the computer of Figure 2.
• an environment 10 comprising a plurality of zones Zi in which mobile entities, such as pedestrians, are likely to move.
• the mobile entities will be considered to be pedestrians without loss of generality.
• the elements described below are indeed applicable to other mobile entities such as cyclists, motorcyclists, car drivers, robots, drones, or even animals.
• environment 10 is a station hall or a metro station.
• the environment 10 comprises thirteen zones respectively named Zi to Zi 3 .
• the movement of pedestrians in the zones Zi is for example free.
• the pedestrians are able to move from any first zone Zi to each adjacent second zone Zj.
• the zones Zi, Z 3 , Z 7 correspond to entrances to the environment 10.
• the zones Z 5 and Z 6 correspond to train or metro platforms through which pedestrians are forced to pass to enter the train or metro. It is clear that areas of congestion are likely to appear.
• the environment 10 further comprises at least one sensor capable of obtaining information on the pedestrians in the environment 10, and preferably in the zones Zi.
• the at least one sensor comprises a plurality of cameras, each pointed towards one or more zones Zi and capable of obtaining images of pedestrians in said zone or zones Zi.
• Computer 20 is preferably a computer.
• the computer 20 is an electronic computer suitable for manipulating and/or transforming data represented as electronic or physical quantities in registers of the computer 20 and/or memories, into other similar data corresponding to physical data in memories, registers or other types of display, transmission or storage devices.
• the computer 20 includes a display unit 22 and a data processing unit 24.
• the display unit 22 is for example a computer monitor connected to the processing unit 24.
• the display unit 22 is suitable for displaying, intended for a user, information coming from the unit treatment 24.
• the processing unit 24 is connected to the display unit 22 and comprises a processor 26 and a memory 28.
• the processing unit 24 is for example connected to the sensor(s) described above.
• the memory 28 preferably stores a computer program product 30 comprising software instructions which, when executed by the processor 26, implement a method 100 for characterizing the movement of pedestrians in the environment 10.
• the computer program product is stored on an information medium, not shown.
• the information medium is a medium readable by the computer 20, usually by the processing unit 24.
• the readable information medium is a medium suitable for storing electronic instructions and capable of being coupled to a bus of a computer system.
• the information medium is a diskette or a floppy disk (from the English name "floppy disk"), an optical disk, a CD-ROM, a magneto-optical disk, a ROM memory, a memory RAM, an EPROM memory, an EEPROM memory, a magnetic card, an optical card or a USB key.
• FIG. s representing a flowchart of the different phases of the method 100 for characterizing the movement of pedestrians.
• the method 100 includes a learning phase 110 and an exploitation phase 120.
• the learning phase 110 is for example implemented prior to an implementation of the exploitation phase 120 in the environment 10.
• the learning phase 110 aims to carry out machine learning of a model from so-called labeled data which are pairs of data each comprising model input data and corresponding outputs.
• the purpose of learning is therefore to modify the model so that it provides, when it receives the input data of a pair of data, the output data associated in the labeled data.
• the learning phase 110 includes a first step 121 of receiving an adjacency matrix A indicating the possible displacements between two adjacent zones.
• the adjacency matrix A indicates for each zone Zi the other zones Zj which are adjacent to it, that is to say contiguous.
• the zones Zi are numbered and the adjacency matrix A is a square matrix comprising for each zone Zi, a row and a column.
• the adjacency matrix A includes a coefficient Aj equal to one if the zones Zi and Zj are adjacent, or equal to zero if the zones Zi and Zj are not adjacent .
• the adjacency matrix A is as follows:
• the adjacency matrix A is then a symmetric matrix since, if a first zone Zi is adjacent to a second zone Zj, then the second zone Zj is also adjacent to the first zone Zi.
• the movements between the zones are constrained.
• the pedestrians are able to move towards only part of the zones Zj adjacent to said zone Zi.
• the adjacency matrix A is not symmetric.
• the coefficient Aij corresponding to the passage from a first zone Zi located before customs, to a second zone Zj located after customs is equal to one.
• the coefficient Aj corresponding to the reverse passage is equal to zero.
• the learning phase 110 includes a second step 122 of receiving a database.
• the database comprises, for each of a plurality of successive instants, Tk a pair of data.
• Each pair of data comprises on the one hand a number of pedestrian(s) in each zone Zi, and on the other hand a set of parameters characterizing the movement of the pedestrians in the environment 10.
• the numbers of pedestrian(s) in each zone are for example grouped in the form of a vector.
• the set of parameters forms for example a vector.
• the set of parameters comprises for example one or more parameters chosen from a group consisting of: an average speed of movement of pedestrians in the environment 10, a maximum flow of pedestrians at one or more points of the environment 10, an average flow of pedestrians at one or more points of the environment 10, a frequency of arrival near the environment 10, of transport vehicle(s), and for at least a first and a second point of the environment, a proportion of the pedestrians at the second point who have previously passed through the first point, and for at least a first point of the environment, a second point of the environment and at least one path between the first and second points, a proportion of pedestrians using said path among the pedestrians going from the first to the second point.
• the learning phase 110 includes a step 123 of training a model as a function of the database and the adjacency matrix A.
• the model is configured to provide, from a number of pedestrians (s) in each zone Zi, a respective set of parameters characterizing the movement of pedestrians in the environment 10.
• the model is for example a neural network.
• the neural network comprises an input layer comprising a neuron for each zone Zi of the environment 10, one or more hidden layers comprising a plurality of neurons each, and an output layer comprising a neuron for each parameter of the parameter set.
• the neural network includes connections between each neuron of one layer to a neuron of a next layer and, for each connection, an adjustable weight.
• the model is preferably a Graph Neural Network or a Graph Convolutional Network.
• Such networks include a plurality of hidden layers each comprising as many neurons as the input layer.
• these networks are suitable for receiving the adjacency matrix A and for including connections only for the possible displacements according to the adjacency matrix A.
• GNN and GCN networks include a lower number of connections than a classic deep neural network, and consequently fewer adjustable weights, thus making it possible to accelerate the learning of the neural network and to stabilize the learning. of the neural network.
• the training step 123 optionally comprises a first calculation sub-step 124, for each zone Zi, of an average number of pedestrian(s) in the zone Zi.
• the average number of pedestrian(s) is for example calculated over a sliding time window.
• the sliding window comprises for example a predefined number N of successive instants Tk.
• the window slippery is applied to the number of pedestrian(s) of the data pairs in the database.
• the predefined number N of successive instants Tk is for example equal to five, ten, fifteen, or twenty.
• the training step 123 includes a second calculation sub-step 125 of at least one set of parameters averaged over the sliding time window.
• the calculation of the or each set of averaged parameters is analogous to that of the average number of pedestrian(s) in each zone Zi.
• the same sliding time windows are applied for the calculation of the average number of pedestrian(s) in each zone Zi during the first calculation sub-step 124 and for the calculation of the sets of average parameters during the second sub-step calculation 125.
• the first 124 and second 125 calculation sub-steps are carried out simultaneously by the processing unit 24.
• the training step 123 advantageously comprises a sub-step 126 of normalization of the average number of pedestrian(s) in each zone Z i; to obtain a normalized vector VN whose algebraic norm is preferably equal to one. Normalization sub-step 126 is also referred to as standardization sub-step.
• T is the transpose operator, is the square root function
• Vi is the i-th component of vector V.
• the average numbers of pedestrian(s) are for example grouped together in an average vector VM.
• the average vector VM is then normalized, to obtain a normalized vector VN for example according to the following equation:
• VN - VM iim
• the training step optionally further comprises a substep of application 127, to the model, of a supervised learning algorithm from the or each normalized vector VN, from the or from each set of averaged parameters and from the adjacency matrix A.
• the supervised learning algorithm is for example a stochastic gradient descent algorithm.
• the supervised learning algorithm modifies the weights of the model so that, when a normalized vector VN is supplied as input to the model, the model supplies as its output a set of parameters as close as possible to the averaged set of parameters corresponding to said normalized vector VN.
• the adjacency matrix A makes it possible to limit the number of weights to be modified by the supervised learning algorithm.
• the model is said to be trained and the learning phase 110 ends.
• the exploitation phase 120 is for example implemented following the learning phase 110, or several hours, days, or months after the learning phase 110.
• the exploitation phase 120 includes a third reception step 131 during which, for each zone Z i; a number of pedestrian(s) in zone Zi is received.
• the third reception step 131 comprises a sub-step 132 of obtaining information(s) from the at least one sensor present in the environment 10.
• the at least one sensor includes cameras
• the images captured by the cameras are obtained by the processing unit 24.
• the third reception step 131 optionally comprises a sub-step 133 of determining a number of pedestrian(s) in each zone Zi by processing the information obtained during the obtaining sub-step 132.
• the determination of the number of pedestrian(s) in each zone Zi is for example carried out by applying an algorithm of image processing known per se.
• Such an algorithm is based for example on artificial intelligence tools such as models resulting from machine learning.
• the images are then VGA (Video Content Analysis) data.
• the exploitation phase 120 comprises a step 134 of application, to the number of pedestrian(s) received for each zone Zi, of the model trained during the learning phase 110, to obtain a respective set of parameters characterizing the movement pedestrians in the environment 10.
• the set of parameters obtained during application step 134 then depends on the number of pedestrian(s) received.
• the exploitation phase 120 includes a first step 135 of detecting an anomaly in the set of parameters.
• the anomaly is for example the fact that one of the parameters of the parameter set exceeds a predefined threshold.
• the set of parameters includes the average speed of pedestrians in environment 10
• the average speed is greater than a first predefined value such as 10km/h
• an anomaly is detected indicating that the pedestrians are running instead of to walk.
• Such an anomaly is for example characteristic of an abnormal crowd movement.
• the set of parameters includes the maximum flow of pedestrians at a point in the environment 10, if this flow is less than a third predefined value such as 100 pedestrians per minute, then an anomaly is detected indicating that an element obstructs the aforementioned point of the environment 10.
• the set of parameters comprises the frequency of arrival near the environment 10, of transport vehicles
• this frequency is lower than a fourth predefined value
• an anomaly is detected indicating a problem of movement of transport vehicles.
• Such a problem is for example likely to create congestion in the environment 10 because the pedestrians risk not being extracted from the environment 10 quickly enough by the transport vehicles.
• predefined threshold(s) are likely to change over time, and in particular over the course of the day or the year.
• the exploitation phase 120 includes, if an anomaly is detected during the first determination step 135, a first step 136 of implementing a corrective action in the environment 10. It is clear that the corrective action is an action performed automatically or an action recommended for its implementation by a user, in order to overcome the anomaly detected.
• the corrective action is, for example, an opening of a new zone Zi in the environment 10, an extension of an already existing zone Zi or an opening of a passage between two zones Zi not yet adjacent.
• the first implementation step 136 includes for example the transmission of an alert to an operator, via the display unit 22. The operator is then able to implement one of the actions of aforementioned corrections.
• the first implementation step 136 includes, for example, sending an actuation command to the environment 10 to perform one of the aforementioned corrective actions, autonomously.
• the exploitation phase 120 comprises a step 137 of predicting the movement of pedestrians in the environment 10.
• the prediction step 137 comprises for example the simulation of the movement of the pedestrians in the environment 10, by a simulator included in the memory 28 of the processing unit 24, from the set of parameters obtained by applying the trained model to the number of pedestrian(s) in each zone Zi.
• the parameters of the set of parameters are input quantities of the simulator making it possible to predict, as faithfully as possible, the evolution of the movement of the pedestrians in the environment 10.
• these parameters are calibration parameters of the simulator.
• a predicted scenario is determined.
• the predicted scenario describes the evolution of the movement of pedestrians in the environment 10 during a predefined duration.
• the predefined duration is for example equal to five minutes, ten minutes, thirty minutes, one hour, two hours, or five hours.
• the predicted scenario includes an evolution of quantities simulated by the simulator.
• the simulated quantities are for example the position of each pedestrian in the environment 10, or the density of pedestrians in each zone Zi
• the exploitation phase 120 advantageously comprises a second step 138 of detecting a predicted anomaly based on the predicted scenario.
• the second detection step 138 is analogous to the first detection step 135 except that the predicted anomaly is detected in the simulated quantities rather than in the parameters of the set of parameters.
• the exploitation phase 120 advantageously comprises a second implementation step 139 analogous to the first implementation step 136.
• the corrective action is implemented in the environment 10 in a preventive manner.
• the corrective action(s) implemented during the second implementation step 139 are intended to prevent the detected predicted abnormality(ies) from taking place in the environment 10.
• the exploitation phase 120 is for example reiterated with the reception 131 of new numbers of pedestrian(s) in the zones Zi.
• the method according to the invention makes it possible to characterize the movement of pedestrians in the environment 10 in a simple and realistic manner.
• a temporal dynamic of the pedestrians is taken into account.
• taking into account several pairs of data for each input of the learning of the model makes it possible to account for a temporal evolution of the movement of pedestrians.
• the learning of the model is accelerated.
• the learning of the model from the adjacency matrix A is of better quality than without the use of the adjacency matrix A.
• the use of the adjacency matrix A makes it possible to account for the spatiality of the environment 10 and takes advantage of the knowledge of the distribution of the zones Zi in the environment 10.
• the optional sub-steps of obtaining 132 and determining 133 make it possible to receive the number of pedestrian(s) in each zone Zi in a simple and substantially inexpensive way.
• the optional prediction step 137 in combination with the application step 134 makes it possible to faithfully account for an evolution of the movement of pedestrians in the environment 10 and therefore to predict potential future anomalies.
• detection 135, 138 and implementation 136, 139 make it possible to ensure the correct movement of pedestrians in the environment 10.

Landscapes

• Engineering & Computer Science (AREA)
• Physics & Mathematics (AREA)
• Theoretical Computer Science (AREA)
• Evolutionary Computation (AREA)
• Geometry (AREA)
• Artificial Intelligence (AREA)
• Medical Informatics (AREA)
• Computer Hardware Design (AREA)
• Computer Vision & Pattern Recognition (AREA)
• General Engineering & Computer Science (AREA)
• General Physics & Mathematics (AREA)
• Software Systems (AREA)
• Traffic Control Systems (AREA)
• Machine Translation (AREA)
• Character Discrimination (AREA)
• Image Analysis (AREA)
• Electrically Operated Instructional Devices (AREA)
• Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Applications Claiming Priority (2)

Publications (1)

Family

ID=82319614

Family Applications (1)

Country Status (3)

• 2021
  • 2021-12-27 FR FR2114547A patent/FR3131407B1/fr active Active
• 2022
  • 2022-12-27 WO PCT/EP2022/087919 patent/WO2023126420A1/fr not_active Ceased
  • 2022-12-27 EP EP22839386.4A patent/EP4457682A1/de active Pending

Also Published As

Similar Documents

Legal Events