System and method for controlling autonomous vehicle using probabilistic occupancy grid
By using reference map data and sensor data to update the assumption probability scores of obstacles in the occupied grid in autonomous vehicles, the risk of vehicle control under environmental uncertainty is solved, safer and more efficient vehicle decisions are achieved, and computing resources are moderately saved .
Patent Information
- Application Number
- CN202280094732.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-13
- Publication Date
- 2024-11-19
AI Technical Summary
When controlling autonomous vehicles in an incompletely known environment, prior art is difficult to balance the impact of old and new measurement data, resulting in vehicle control decisions that may be based on outdated data, increasing risks, and the binary definition of traditional occupancy grids is not sufficient to express complex environments Obstacles in assumptions of possibility.
By initializing the occupancy grid with reference map data, updating the probability scores of the obstacle hypothesis in combination with sensor data, and controlling autonomous vehicles based on these probability score evolutions, the Bayesian rule and scoring model are used to update and aggregate the probability scores, gradually approaching The limit value reflects the aging of the measurement data and reduces the consumption of computing resources.
Effectively balance the impact of new and old measurement data, reduce the risk of outdated data, improve the safety and computing efficiency of vehicle control, and moderately retain computing resources for easier processing of other tasks.
Smart Images

Figure CN118984980A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of autonomous vehicle control, and in particular, it proposes a system and method for controlling an autonomous vehicle using a probabilistic occupancy grid. Background Art
[0002] Occupancy grids are a common tool for controlling unmanned vehicles or robots moving in an imperfectly understood environment. Conceptually, the surface (ground, road, track, land, floor, etc.) on which the vehicle moves is discretized into cells, which are labeled as 'occupied' or 'free'. In the field of autonomous vehicles, several different kinds and definitions of occupancy grids coexist. For example, an occupancy grid can be a mapping from grid cells to corresponding maneuverability scores. Different maneuverability score values can indicate that there is an obstacle in the grid cell, that the grid cell can be safely driven, or that the exact maneuverability is uncertain. Alternatively, an occupancy grid can be a mapping from grid cells to random variables indicating the presence of obstacles at those locations. In addition, the user community has accumulated a lot of experience with occupancy grids, and there are now very efficient auxiliary software functions, including geometry search algorithms.
[0003] For a general introduction to occupancy grids, see Probabilistic Robotics (Series on Intelligent Robots and Autonomous Agents), S. Thrun et al., MIT Press, 2005, ISBN 9780262201629.
[0004] The occupancy grid is typically initialized using map data, which may be provided by a Geographic Information System (GIS). It is not uncommon to have use cases where the GIS data or other initialization data in the occupancy grid needs to be updated after vehicle control has begun. The update may be point-by-point (e.g., at a scheduled or triggered refresh event) or continuous, and it may be based on data from fixed or onboard sensors such as radar, lidar, optical and acoustic sensors. In such use cases, the occupancy grid has repeatedly proven to be a suitable medium for absorbing such subsequent data even from multiple data sources and integrating it with the information already possessed by the occupancy grid. Sensor fusion and SLAM (Simultaneous Localization and Mapping) are technologies related to the process of integrating and combining data.
[0005] WO2021097409A1 discloses a computer-based method for generating artificial sensor data suitable for training an autonomous vehicle before it is placed in a real traffic environment for operation. The disclosed method uses sensor data collected when the autonomous vehicle passes through a scene of interest one or more times, such as optimal camera and lidar data. In this method, a grid is generated for each facet (surface element) in a facet map. The grid can have multiple grid cells, for example, a square grid centered on the point centroid of the facet. A generative model implemented as a generative adversarial network (facet-GAN) is used to convert facet rendering into a more realistic image. After applying the facet-GAN, gaps, holes and other spaces missing in the first image are automatically filled. Other objects (including roads, traffic lights, lamp posts and signs) become clearer and more realistic. Shape and size parameters are used to specify the shape and size of the facet, and its orientation can be represented by a normal vector. Summary of the invention
[0006] One object of the present disclosure is to provide techniques for controlling autonomous vehicles in an incompletely known environment explored by sensors. Another object is to balance the impact of new measurements from sensors and relatively late measurements in the vehicle control process. Another object is to reduce the risk that vehicle control decisions are based on outdated measurements, and the risk that they become uncertain or dangerous as a result. Another object is to implement such vehicle control techniques based on occupancy grids. Yet another object is to develop the conventional binary definition of occupancy states in occupancy grids ('occupied' or 'free') into the possibility of formulating and evaluating multiple parallel hypotheses about specific grid cells. Finally, an overall object of the present disclosure is to propose control techniques with moderate computational requirements, i.e., techniques that can be expected to contribute a limited amount of processor load, thereby reserving computational resources for use by other tasks on the same system.
[0007] At least some of these objects are achieved by the invention as defined by the independent claims. The dependent claims relate to advantageous embodiments of the invention.
[0008] In a first aspect, a method of controlling at least one autonomous vehicle movable on a surface is provided. The method comprises: initializing an occupancy grid using reference map data, wherein the occupancy grid comprises a plurality of cells, each cell representing a portion of the surface and associated with a plurality of obstacle hypotheses and corresponding probability scores; repeatedly acquiring measurement data from one or more sensors; associating the acquired measurement data with corresponding cells in the occupancy grid; in each of the corresponding cells, updating a probability score of at least one of the obstacle hypotheses based on the acquired measurement data; outside of the corresponding cells, evolving the probability scores based on a time elapsed since the probability scores were last updated; and controlling the autonomous vehicle based on the occupancy grid.
[0009] In the present disclosure, "obstacle" should be broadly understood as anything that may impair or prevent normal movement of the vehicle (e.g., potholes, snow, ice, contaminants, obstructing physical objects, land, missing road, missing road material within the nominal range of the road surface). A "probability score" can be any quantity that provides sufficient guidance for the autonomous vehicle control to be performed in the current use case. The probability score can have a maximum value (e.g., corresponding to a statement in natural language such as 'the hypothesis is confirmed'), a minimum value (e.g., 'the hypothesis is refuted'), and a neutral value (e.g., 'the hypothesis is uncertain'). The probability score can be represented as a discrete or continuous numeric variable. Specifically, the probability score can be defined as having a value within the real number interval [0,1]; however, even with such a definition, the probability score does not necessarily correspond faithfully to a probability in the statistical sense. On this basis, it should be understood that the hypothesis "the grid cell contains a usable road" or another favorable hypothesis can be expressed as a mirrored unfavorable hypothesis "the road is missing". Then, if the vehicle controller requires that the probability score of the favorable hypothesis is at least (For a configurable constant ), this is equivalent to requiring that the probability score of the unfavorable hypothesis is at most .
[0010] A particularly advantageous property of the method according to the first aspect is that it is able to balance the impact of new measurement data from the sensor and relatively late measurement data. This is achieved by the fact that each cell in the occupancy grid undergoes a probability score update or probability score evolution when new measurement data is acquired. The evolution of the probability score may reflect the aging of the measurement data used to last update the probability score. In practical implementations, there are at least two ways to determine whether new measurement data has been acquired:
[0011] (a) Define the length of the and after expiration of each time period, checking which units are associated with the measurement data; or
[0012] (b) After each repetitive step of acquiring measurement data (e.g., polling available sensors), it is checked which cells are associated with the measurement data.
[0013] Option (b) can be implemented regardless of whether the step of repeatedly acquiring measurement data is performed periodically or in an event-triggered manner. Specifically, the method can be implemented so that at least some of the step of acquiring measurement data and subsequent steps are repeated periodically or in an event-triggered manner.
[0014] In some embodiments, the probability score evolves by gradually approaching a predefined limit value. The limit value may, for example, represent an unconfirmed obstacle hypothesis (a neutral value, such as 0.5). The vehicle controller may treat grid cells with unconfirmed obstacle hypotheses as uncertain, and its decision making will tend to be cautious in order to ensure safety.
[0015] In a further development of these embodiments, the first predefined limit value applies to cells outside the field of view of the sensor from which the measurement data is acquired, and the second predefined limit value applies to cells in the field of view. The first predefined limit value may represent an unconfirmed obstacle hypothesis (e.g., a neutral value such as 0.5), and the second predefined limit value may represent a refuted obstacle hypothesis (e.g., a minimum value such as 0.0).
[0016] In some embodiments, the probability score is evolved in any intermediate unit between the corresponding unit and the corresponding sensor location unit. In the terminology of the present disclosure:
[0017] - The measurement data consist of points in space where reflection, scattering and / or diffraction have been recorded by one or more sensors;
[0018] - The corresponding unit can be calculated as the projection of the point in space onto the occupancy grid (e.g., the spatial coordinates of the measurement data are The coordinates are discretized so that they correspond to grid cells and the coordinate);
[0019] - the sensor location cell is the projection of the sensor location on the occupancy grid; and
[0020] - An intermediate unit is understood to be a unit on the line connecting one of the corresponding units with the sensor location unit of the active sensor.
[0021] Therefore, the intermediate cells are not part of the corresponding cell, which can lead to a significant saving of computational resources. In fact, a large number of grid cells undergo computationally cheaper probability score updates rather than probability score evolutions.
[0022] In some embodiments, the updating of the probability scores in the corresponding cells (i.e., those cells that have been associated with the measurement data) is based on a comparison of the measurement data with the reference map data. The comparison may be configured to determine whether the measurement data confirms the obstacle hypothesis (or at least is consistent with the obstacle hypothesis), in which case the probability score is increased. The comparison may involve visual features or other physical and chemical quantities recorded by the sensor.
[0023] In particular, such a comparison of the measurement data with the reference map data may involve a measured altitude according to the measurement data and an altitude according to the reference map data. A situation where the measured altitude is equal to the altitude according to the reference map data may be understood as refuting (i.e., not supporting) any obstacle hypothesis for the road indicated by the reference map. A situation where the altitude is equal may be understood as supporting a surface-type obstacle hypothesis (e.g., ice, liquid contaminants). A situation where the measured altitude is greater than the altitude according to the reference map data may be understood as proving an object-type obstacle hypothesis (e.g., the presence of a physical object). A situation where the measured altitude is less than the altitude according to the reference map data may be understood as confirming a hole-type obstacle hypothesis (e.g., the presence of potholes, missing road material). Experiments and simulations have shown that the altitude is an accurate and computationally streamlined metric that ensures rapid convergence to a reliable occupancy grid.
[0024] In some embodiments, the updating of the probability score includes applying a recursive rule. The recursive rule may be, for example, a version of the Bayesian rule.
[0025] In some embodiments, the probability scores are updated according to a predefined scoring model. The scoring model may, for example, reflect the reliability and accuracy of the sensors used; if the sensors are prone to errors and / or the observation conditions are poor, the support will be relatively weak. The scoring model may depend on the (time) update rate of the sensors and / or the spatial measurement density. Furthermore, the scoring model may be specific to the corresponding obstacle hypothesis and / or to the sensor from which the measurement data is obtained. In embodiments utilizing the Bayesian rule, the evaluation of the rule may involve various aspects of the scoring model.
[0026] In some embodiments, the step of controlling the autonomous vehicle includes aggregating the cell-by-cell probability scores of the obstacle hypotheses into an occupancy probability. As with the probability scores, the occupancy probability need not correspond faithfully to a probability in a statistical sense. It can be used to evaluate according to a configurable threshold that represents the degree of certainty required for an operator to consider a cell as a drivable cell and that will confirm that it is safe to plan a vehicle trajectory through the cell. In a further development of such embodiments, the aggregation of the cell-by-cell probability scores is restricted to cells on the tentative trajectory of the vehicle or cells in the vicinity of the vehicle.
[0027] In a second aspect, a vehicle controller is provided, the vehicle controller being configured to control at least one autonomous vehicle movable on a surface. The autonomous vehicle is controlled based on an occupancy grid comprising a plurality of cells, each cell representing a portion of the surface and associated with a plurality of obstacle hypotheses and corresponding probability scores. The vehicle controller comprises: an input interface for receiving measurements from one or more sensors; processing circuitry; and an output interface for providing commands to the autonomous vehicle. The processing circuitry of the vehicle controller is configured to perform the method of the first aspect.
[0028] The invention also relates to a computer program comprising instructions for causing a computer or in particular a controller to perform the above method. The computer program may be stored or distributed on a data carrier. As used herein, a "data carrier" may be a temporary data carrier, such as a modulated electromagnetic wave or light wave, or a non-temporary data carrier. Non-temporary data carriers include volatile and non-volatile memories, such as permanent and non-permanent storage media of the magnetic, optical or solid-state type. Still within the scope of a "data carrier", such memory may be fixedly mounted or portable.
[0029] Generally, unless otherwise explicitly defined herein, all terms used in the claims should be interpreted according to their ordinary meaning in the art. Unless otherwise explicitly stated, all references to "a / an / the element, device, component, means, step, etc." will be interpreted openly as referring to at least one instance of the element, device, component, means, step, etc. Unless explicitly stated otherwise, the steps of any method disclosed herein do not have to be performed in the exact order described. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Various aspects and embodiments will now be described by way of example with reference to the accompanying drawings, in which:
[0031] Figure 1 is a flow chart of a method of controlling an autonomous vehicle according to embodiments herein;
[0032] Figure 2 is a perspective view of a surface on which a vehicle can move and in which different types of obstacles are present;
[0033] Figure 3 is a block diagram of a vehicle controller according to embodiments herein that receives measurement data from sensors and generates control signals to the vehicle;
[0034] Figure 4 shows the time evolution of the probability score of the obstacle hypothesis associated with a grid cell, where the probability score is close to the value 0.5 in the absence of new measurement data;
[0035] Figure 5 shows the time evolution of the probability score of the obstacle hypothesis associated with a grid cell, where the probability score approaches the value 0.0 in the absence of new measurement data;
[0036] Figure 6 Various geometric features of sensor-captured measurement data and resulting updates of probability score values in some grid cells are shown in the form of a perspective view of a surface on which a vehicle may move;
[0037] Figure 7 is a perspective view of an example occupancy grid initialized using reference map data, where the cell heights correspond to the altitude and the grouping of cells into so-called tiles is visible;
[0038] Figure 8 Corresponds to Figure 7 , covered by another layer, showing the updated probability score of an obstacle hypothesis;
[0039] Fig. 9 Corresponds to Figure 7 , partially covered by a layer showing a snapshot of all current probability scores (i.e., updated and evolving probability scores) for one obstacle hypothesis; and
[0040] Fig.10 Corresponds to Figure 7 , partially covered by a layer, showing a snapshot of the occupancy probabilities based on the aggregated probability scores of all obstacle hypotheses. DETAILED DESCRIPTION
[0041] Various aspects of the present disclosure will now be described more fully below with reference to the accompanying drawings, in which certain embodiments of the present invention are shown. However, these aspects may be embodied in many different forms and should not be construed as limiting; rather, these embodiments are provided by way of example so that the present disclosure will be thorough and complete, and the scope of all aspects of the present invention will be fully conveyed to those skilled in the art. Throughout this specification, the same reference numerals refer to the same elements.
[0042] Figure 1 1 is a flow chart of a method 100 for controlling at least one autonomous vehicle 210 movable on a surface. The steps of the method 100 to be described below may be performed by a general purpose computer equipped with a suitable communication interface through which the computer can exchange data with sensors and actuators. Specifically, the method 100 may be performed in Figure 3 More specifically, the vehicle controller 300 includes an input interface 310 for receiving measurements from one or more sensors 250, an output interface 330 for providing commands to the autonomous vehicle 210, and a processing circuit 320 for providing commands based on the measurements. Figure 3In the illustrated embodiment, the vehicle controller 300 also includes an optional operator interface 340 that is configured for human-computer interaction; the operator interface 340 can be used for system setup, system maintenance, troubleshooting, manual control of the overriding vehicle 210, and similar purposes. In further embodiments, the processing circuit 320 is configured to delegate some of its computations to a networked (or cloud) processing resource, wherein the vehicle controller 300 can exchange computational tasks and resulting outputs with the networked processing resource via a wireless or wired communication interface.
[0043] like Figure 2 As shown, the sensor 250 providing the measurement data is preferably arranged on, near or above the surface 220 on which the autonomous vehicle 210 moves. The sensor 250 may be fixed (with Figure 2 ) or on-board (as with sensor 250 disposed on autonomous vehicle 210). In addition, each sensor 250 may have a variable or fixed orientation; a viewing angle suitable for determining the presence or absence of an obstacle may include a direction substantially parallel to surface 220, a direction orthogonal to surface 220, or an oblique direction between these directions. Sensor 250 may be configured to apply at least one of the following measurement principles: optics, electromagnetic reflection, electromagnetic scattering, electromagnetic diffraction, lidar, color depth sensing (e.g., RGB-D), millimeter wave radar, ultra-wideband radar. When method 100 is to be performed, it is assumed that sensor 250 has been properly calibrated. In some embodiments, poor calibration or fault conditions may be discovered and corrected by performing the so-called altitude condition described below.
[0044] It is important to recall that in the context of this disclosure, "obstacles" should be broadly understood as anything that may impair or prevent normal vehicle movement, and they may correspond to object-type, surface-type, and hole-type obstacle hypotheses. Figure 2 In the figure, the following obstacles are visible: physical objects 240, liquid contaminants 241, potholes 242. Further examples of obstacles include snow, ice, solid or powdery contaminants, land, missing road, missing road material within the nominal range of the road surface. Autonomous vehicle 210 can be a road vehicle, in particular a truck, bus or construction equipment vehicle. It can be an integral vehicle or an articulated vehicle, or a vehicle combination of two or more units.
[0045] Figure 2 Also included is an external view of the vehicle controller 300, which was previously referenced Figure 3 was discussed. Figure 2, the input interface 310 and the output interface 330 of the vehicle controller 300 are implemented as a single communication interface, which is arranged within the coverage of the corresponding wireless interface 211 installed on the vehicle 210 to be controlled and within the coverage of other wireless interfaces (not shown) on the sensors 250. It should be noted that at least some of the sensors 250 may alternatively provide their measurement data to the vehicle controller 300 via a wired link.
[0046] Back to Figure 1 The execution flow shown in the method 100 First step 110 The method includes initializing an occupancy grid 230, which includes a plurality of cells 231, each cell representing a portion of a surface. The occupancy grid 230 is initialized using reference map data, and the cells are associated with a plurality of obstacle hypotheses and corresponding probability scores.
[0047] Figure 2 2, the division of the surface 220 into a plurality of cells 231 is indicated symbolically, which represent respective portions of the surface and which together form an occupancy grid 230. The granularity of the occupancy grid 230 is not essential to the present invention, but may vary between different embodiments. For road vehicles, the cell area may be approximately 0.01 m 2 , 0.1 m 2 or 1.0 m 2 Cell shapes that are easy to pack homogeneously, such as hexagons and rectangles, are preferred. Areas that are not drivable from the outset (e.g., water areas) need not be included in the occupancy grid 230 at all. Alternatively, non-drivable areas may be included in the occupancy grid 230 and may be identified by an assigned altitude value ( coordinates) are missing.
[0048] Grid cells can be indexed by a pair of integers. To identify, the index and the Cartesian coordinates of the unit It is related by the following discretization function:
[0049]
[0050] The constant is the corresponding segmentation granularity, and Represents rounding to the nearest lower integer (floor function). Discretization and The Cartesian coordinates of a grid cell may refer to a predefined reference point in the cell (e.g., center, northwest corner). Alternatively, a discretization is used to define the grid cell, i.e., the grid cell Should be surface The set of all points where the discretized function maps to the same integer index:
[0051] .
[0052] Processing circuitry 320 of vehicle controller 300 may be configured to maintain a data structure representing occupancy grid 230. In addition to geometric indications about cells 231, the data structure may also include reference map data. Inputting the reference map data into the data structure may be referred to as initialization. The reference map data may include the location of nominally drivable portions of surface 220 (e.g., roads, parking lots). The reference map data may also include terrain indications; in embodiments where probability scores are to be updated based on altitude comparisons, the terrain indications include altitude ( coordinates). Some authors refer to occupancy grids with these characteristics as 2.5-dimensional (2.5D) height grids. Map data can be obtained from commercial geographic information systems (GIS) or from open data sources such as the OpenStreetMap Foundation in Cambridge, UK.
[0053] Furthermore, in the data structure, each cell 231 of the occupancy grid 230 is associated with a plurality of obstacle hypotheses and corresponding probability scores. Although probability scores may be assigned independently to each cell 231, the occupancy grid 230 may be established in such a way that all cells share a common set of obstacle hypotheses. Alternatively, different cells 231 are associated with different sets of obstacle hypotheses, e.g., when it is a priori obvious that a certain type of obstacle will never appear in a certain area, the corresponding obstacle hypothesis need not be initialized. As initially mentioned, a "probability score" may be any quantity that provides sufficient guidance for the autonomous vehicle control to be performed. A probability score may have a maximum value ('hypothesis confirmed'), a minimum value ('hypothesis refuted'), and a neutral value (e.g., 'hypothesis uncertain'). A probability score may be represented as a discrete or continuous numeric variable. In the examples to be described below, the probability score is defined as having a value in the real number interval [0,1].
[0054] In method 100 Second step 112 In the embodiment of the present invention, measurement data is obtained from one or more sensors 250. Figure 2Suitable types of sensors 250 and their arrangement relative to the surface 220 are discussed. The entity performing the second step 112 may obtain the measurement data by requesting the measurement data from the sensor 250, or may receive the measurement data in a communication sent to it by the sensor 250 periodically or in an event-triggered manner. Yet another option is to use a publish-subscribe (PubSub) arrangement, in which the sensor 250 is configured to store the measurement data in a shared memory. The measurement data may include points in space where reflections, scattering, or diffraction have been recorded by one or more sensors; points may be identified according to their Cartesian coordinates. A collection of such points may be referred to as a point cloud.
[0055] The second step 112 is repeated within a single execution of the method 100. The second step 112 may be repeated at predefined intervals, at predetermined points in time, whenever new measurement data is received from the sensor 250, or in an event-triggered manner. For example, if the quality indicator of the occupancy grid 230 is below a predefined threshold, re-execution of the second step 112 may be triggered. In some embodiments, for each new execution of the second step 112, the subsequent third step 114, fourth step 116, and fifth step 118 (and optionally, sixth step 120) are repeated.
[0056] In method 100 Third Step 114 In , the acquired measurement data are associated with the corresponding cells in the occupancy grid. Figure 6 An example of such an association is shown, where L1 is a line from sensor 250 to the point on obstacle 240 where reflection, scattering or diffraction occurs. Figure 6 The dashed line in the middle is the vertical line containing the reflection / scattering / diffraction points ( The third step 114 may include using the discretization function defined above Mapping reflection / scattering / diffraction points in measurement data , the discretization function returns a pair of integer indices identifying the corresponding cell 231-5 The cell containing the vertical projection of the sensor 250 is referred to as the sensor location cell 231-1. The cells intersecting the line L2 of the corresponding vertical projection connecting the sensor location and the reflection / scattering / diffraction point are referred to as intermediate cells 231-2, 231-3, 231-4. Alternatively, the intermediate cells may be defined as all cells of the occupancy grid intersecting the line connecting the sensor location cell and any of the corresponding cells.
[0057] In the case of these corresponding units ( Figure 6 Each of 231-5) in Fourth Step 116In some embodiments of method 100, the probability scores of at least one of the obstacle hypotheses are updated according to the acquired measurement data. In some embodiments of method 100, the probability scores are updated according to a predefined scoring model. The scoring model may, for example, reflect the reliability and accuracy of the sensor 250 used; if the sensor is prone to errors and / or the observation conditions are poor, the impact of the new measurement data will have a relatively weak effect on the probability score. The scoring model may depend on the (time) update rate and / or the (spatial) measurement density of the sensor. In addition, the scoring model may be specific to the corresponding obstacle hypothesis and / or to the sensor 250 from which the measurement data is acquired. In order to establish a new scoring model, the implementer may rely on theoretical reasoning or early experience to initialize the scoring model with numerical values (ansatz). Subsequently, the model may be refined and improved from a structural or quantitative perspective based on simulations or experiments.
[0058] In some embodiments of method 100, the updating of the probability score includes applying 116.1 a recursive rule. The recursive rule may be, for example, a version of the Bayesian rule:
[0059]
[0060] in, is the obstacle hypothesis before the update The probability score of (i.e., the prior probability, the initial confidence level in the obstacle hypothesis), is the probability score after including the measured data, and the quotient Can be interpreted as measurement data Assume that obstacles are Thus, the scoring model can include the quotient of each obstacle hypothesis. The numerical value of .
[0061] Specifically, the probability score may be updated based on the altitude measured by one of the sensors 250. The altitude indicated by the reference map data More precisely, the probability score is updated depending on the terrain condition or the altitude condition. The terrain condition ensures that the measured data roughly matches the obstacle hypothesis. Different altitude conditions may be suitable for the hole-type, surface-type or hole-type obstacle hypothesis, as shown in Table 1.
[0062] exist Fifth Step 118(This step is performed for each cell in the occupancy grid 230 except the corresponding cell), the probability score is evolved based on the time that has passed since the last update of the probability score in the cell. (In some embodiments, the fifth step 118 can also be performed for the corresponding cell, which can result in a simpler implementation. It is expected that the update of the probability score will outnumber the effect of the evolution.) Instead of performing the more expensive fourth step 116, the intermediate cells 231-2, 231-3, 231-4 are processed according to the fifth step 118, which can result in a saving in computational effort.
[0063] The evolution of the probability scores may include gradually approaching each probability score to a predefined limit value. In some embodiments of method 100, the first predefined limit value Applicable to cells outside the field of view of the sensor from which the measurement data is acquired, and the second predefined limit value Applicable to cells in the field of view. In such embodiments, the first predefined limit value may represent an unconfirmed obstacle hypothesis: The second predefined limit value may represent a refuted obstacle hypothesis: .
[0064] Figure 4 The probability scores of the obstacle hypotheses occupying the cells of the grid 230 as a function of time are shown. The initial value at is about 0.75. and The new measurement data at is associated (i.e., it becomes the corresponding unit), which is now assigned the values 0.9 and 0.2. to Between and In the following time period, the probability score tends to a limit value of 0.5. This reflects the empirical fact that the credibility of measurement data from a time-varying system decreases as the measurement data ages. The limit value of 0.5 is a typical value for cells outside the field of view of the sensor 250.
[0065] Similarly, Figure 5 The probability scores of different obstacle hypotheses for the same or different cells of the occupancy grid are shown. The probability score is associated with new measurement data at time. In the time period when no new measurement data is received, the probability score tends to the limit value 0.0. The probability score is The limit value 0.0 is characteristic of the cells within the field of view of the sensor 250 .
[0066] If a movable sensor 250 or a sensor 250 with a variable orientation (e.g., a pan-tilt-zoom camera) is used, the probability scores for the same obstacle hypothesis and the same grid cell may tend to different extreme values at different time intervals as a result of the field of view conditions. Figure 4 and Figure 5 2. This situation is different from that depicted in . This situation may occur if sensor 250 is moved or reoriented in such a way that a grid cell enters or leaves the field of view. Whether a grid cell is within the field of view of sensor 250 may be determined by basic geometric calculations based on the position, orientation, and viewing angle of sensor 250. Information about the local elevation of surface 220 may be used to improve this calculation.
[0067] Computationally speaking, the evolution of the probability score may be achieved in any suitable manner. ), the evolution can be performed by applying the same additive or multiplicative constant for each time period. For example, the probability score can be driven toward 0.0 by reducing the probability score by a certain percentage in each time period. The generally applicable evolution equation can be expressed as follows:
[0068]
[0069] in It is assumed and time The probability score at ; is the time when the probability score was last updated; is the applicable limit; and is a monotonically decreasing function, such that and .function may have a linear or nonlinear dependence on time. For example, it may be a decreasing exponential function that tends to 0 in finite time (e.g., ) or an affine function. The above evolution equation has an inherent record of the elapsed time and can therefore be used even if the second step 112 is not repeated regularly.
[0070] Table 1 summarizes the foregoing discussion of the processing of the probability scores in the fourth step 116 and the fifth step 118. The left side relates to updates based on measurement data, while the right side relates to evolution over time. Table 1 relates to embodiments where the updating of the probability scores depends on altitude conditions. If the altitude conditions are not met, then in some embodiments the probabilities may remain unchanged. In other embodiments, if the altitude conditions are not met, then the probability scores may evolve over time - as if no measurement data were received.
[0071]
[0072] In method 100 Sixth step 120 In the example of FIG. 2 , data from the occupancy grid is used to control autonomous vehicle 210. For example, the data may be used to determine an obstacle-free trajectory for vehicle 210. Alternatively, the generated vehicle trajectory may be allowed to pass through certain types of obstacles, but with penalties (e.g., hole-type obstacles that may increase mechanical wear) or with requirements to reduce speed and / or limit maximum curvature (e.g., surface-type obstacles where wheel friction is expected to be poor).
[0073] The data from the occupancy grid may be preprocessed in various ways to make it more suitable for use in controlling the autonomous vehicle 210. For example, the cell-by-cell probability scores for multiple obstacle hypotheses may be aggregated 120.1 into one occupancy probability value for each cell. The occupancy probability can be used to evaluate against a configurable threshold that represents the degree of certainty required for the operator to consider the unit drivable. An example requirement might be ,in is a configurable constant representing the degree of certainty required to consider a grid cell as drivable. The occupancy probability can be obtained by combining the probability scores of all obstacle hypotheses associated with the cell. The combination can be performed based on the simplifying assumption that the obstacle hypotheses are independent of each other. Let Expressing the probability scores of these obstacle hypotheses, a first example criterion for considering a grid cell as a safe driving cell can be stated as
[0074]
[0075] in As mentioned above. The second example criterion can be expressed as
[0076] .
[0077] The left side of the first example criterion can be interpreted as a probability score for the event 'satisfying any single obstacle hypothesis', while the left side of the second example criterion can be interpreted as a probability score for the event 'satisfying at least one obstacle hypothesis'. The second example criterion may be more precise, but may require more processing resources.
[0078] If the aggregation sub-step 120.1 is restricted to cells that the autonomous vehicle 210 may access (e.g., cells located on a tentative trajectory, cells relatively close to the autonomous vehicle 210, and / or cells located substantially in front of the vehicle), the necessary processing load may be limited to some extent.
[0079] As mentioned, the method 100 can be implemented so that the first step 110 of acquiring measurement data and at least some of the subsequent steps are repeated periodically or in an event-triggered manner. The control step 120 can be performed independently of the repetitive cycle of steps 112, 114, 116, 118, the purpose of which is to maintain the data stored in the occupancy grid. As such, the control step 120 can be performed whenever the operator requests the vehicle to move.
[0080] The method 100 has several significant advantages, at least in some embodiments: the measurement data expressed as a point cloud can be used almost without any preprocessing to update the probability score, so that each reflection / scattering / diffraction point has a direct and autonomous impact on the probability score. The measurement data in this form can be calculated by a simplified discretization function Mapped to corresponding cells. The need for expensive point cloud preprocessing (including downsampling, pruning, denoising, and similar operations) is reduced or eliminated. The probability score of each obstacle hypothesis is systematically moved toward (i.e., evolves over time) a certain bias target (i.e., limit value), reflecting model awareness based on missing measurements. Expensive ray tracing can be completely avoided. Probability scores in intermediate cells can be handled by cheap time evolution instead of actual updates.
[0081] Figure 7 1 is a perspective view of an example occupancy grid that has been initialized using reference map data, the occupancy grid being visualized such that the cell heights correspond to the altitude. In addition, the grouping of cells into six so-called "tiles" is visible. The grouping of cells into tiles, while not essential to the present invention, may be used in implementations to facilitate memory management, graphics rendering, and similar housekeeping operations in the processing circuitry performing method 100.
[0082] Figure 8 Corresponds to Figure 7 , overlaid by another layer, shows the updated probability scores for the object-type obstacle hypothesis, where the update is based on the measurement data. Lighter shades correspond to lower probability score values, while darker shades correspond to higher probability score values. The overall brighter, approximately stadium-shaped area corresponds to the field of view of the sensor that delivers the measurement data. Outside this area, all grid cells are processed by evolving the probability scores over time. Inside the stadium-shaped area, the grid cells are selectively processed depending on whether they are corresponding cells (i.e., associated with measurement data), which results in a probability score update or a probability score evolution, respectively.
[0083] Fig. 9 Corresponds to Figure 7 , partially covered by a layer, shows a snapshot of all current probability scores (i.e., updated and evolving probability scores) for the obstacle hypotheses.
[0084] Fig.10 Corresponds to Figure 7 , partially covered by a layer, shows a snapshot based on the aggregated occupancy probability scores. (Conceptually, the aggregated occupancy probability scores can be viewed as a generalized obstacle hypothesis of "the surface is undriveable"). It can be seen that the central area of the partially covered layer can be considered largely unoccupied (no obstacles) with high confidence, while (smaller) obstacles may exist at the near edge, and (larger) obstacles exist at the far edge. Cells towards the left edge of the layer (outside the sensor field of view) have neutral probability score values. For these reasons, no safety conclusions can be drawn about maneuverability.
[0085] Aspects of the disclosure have mainly been described above with reference to a few embodiments. However, as is readily appreciated by a person skilled in the art, other embodiments than the ones disclosed above are equally possible within the scope of the invention, as defined by the appended patent claims.
Claims
1. A method (100) of controlling an autonomous vehicle (210) capable of moving on a surface (220), comprising: Initializing (110) an occupancy grid (230) using reference map data, wherein the occupancy grid comprises a plurality of cells (231), each cell representing a portion of the surface and associated with a plurality of obstacle hypotheses and corresponding probability scores; repeatedly acquiring (112) measurement data from one or more sensors (250); associating the acquired measurement data with corresponding cells in the occupancy grid (114); in each of the corresponding units, updating (116) the probability score of at least one of the obstacle hypotheses based on the acquired measurement data; evolving (118), outside of the corresponding cell, the probability scores based on the time that has passed since the probability scores were last updated (116); and The autonomous vehicle is controlled (120) based on the occupancy grid.
2. The method of claim 1, wherein: Evolving (118) the probability scores includes gradually approaching (118.1) each probability score to a predefined limit value.
3. The method of claim 2, wherein: The first predefined limit value applies to cells outside the field of view of the sensor from which the measurement data is acquired, and the second predefined limit value applies to cells in the field of view.
4. The method of claim 3, wherein: The first predefined limit value represents an unconfirmed obstacle hypothesis.
5. The method according to claim 3 or 4, wherein: The second predefined limit value represents a refuted obstacle hypothesis.
6. A method as claimed in any one of the preceding claims, wherein: The evolving (118) is performed on any intermediate units between the corresponding unit and the corresponding sensor location unit.
7. The method of any of the preceding claims, wherein the updating (116) of the probability score is based on a comparison of the measurement data with the reference map data.
8. The method of claim 7, wherein: The updating (116) of the probability score is based on a comparison of the measured altitude with an altitude indicated by the reference map data.
9. A method as claimed in any one of the preceding claims, wherein: Updating (116) the probability scores includes applying (116.1) a recursive rule, such as Bayes' rule.
10. A method as claimed in any one of the preceding claims, wherein: The probability scores are updated (116) according to a predefined scoring model, the scoring model being specific to the respective obstacle hypothesis and / or specific to the sensor from which the measurement data was acquired.
11. The method of claim 10, wherein: The scoring model depends on the update rate and / or spatial measurement density of the sensor.
12. A method as claimed in any one of the preceding claims, wherein: The obstacle assumptions in the occupancy grid include one or more of: potholes, snow, ice, contaminants, physical objects, land, missing road, missing road material.
13. A method as claimed in any one of the preceding claims, wherein: The sensor applies at least one of the following measurement principles: optics, electromagnetic reflection, electromagnetic scattering, electromagnetic diffraction, lidar, color depth sensing, millimeter wave radar, ultra-wideband radar.
14. A method as claimed in any one of the preceding claims, wherein: The controlling (120) the autonomous vehicle includes aggregating (120.1) the cell-by-cell probability scores of the obstacle hypotheses into occupancy probabilities.
15. The method of claim 14, wherein: The aggregation (120.1) is restricted to cells on the tentative trajectory of the vehicle or cells in the vicinity of the vehicle.
16. A vehicle controller (300) configured to control at least one autonomous vehicle (210) capable of moving on a surface (220), wherein: The autonomous vehicle is controlled based on an occupancy grid (230) comprising a plurality of cells (231), each cell representing a portion of the surface and associated with a plurality of obstacle hypotheses and corresponding probability scores, the vehicle controller comprising: an input interface (310) for receiving measurements from one or more sensors (250); processing circuit (320); and an output interface (330) for providing commands to the autonomous vehicle, Wherein, the processing circuit is configured to perform the method as claimed in any one of the preceding claims.
17. A computer program comprising instructions for causing a vehicle controller as claimed in claim 16 to perform the steps of the method as claimed in any one of claims 1 to 15.
