WO2015181568A2 - Prévision d'emplacements atteignables - Google Patents

Prévision d'emplacements atteignables Download PDF

Info

Publication number
WO2015181568A2
WO2015181568A2 PCT/GB2015/051575 GB2015051575W WO2015181568A2 WO 2015181568 A2 WO2015181568 A2 WO 2015181568A2 GB 2015051575 W GB2015051575 W GB 2015051575W WO 2015181568 A2 WO2015181568 A2 WO 2015181568A2
Authority
WO
WIPO (PCT)
Prior art keywords
vehicle
energy
map
map data
attainable
Prior art date
Application number
PCT/GB2015/051575
Other languages
English (en)
Other versions
WO2015181568A3 (fr
Inventor
Peter ONDRUSKA
Ingmar POSNER
Original Assignee
Isis Innovation Limited
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Isis Innovation Limited filed Critical Isis Innovation Limited
Publication of WO2015181568A2 publication Critical patent/WO2015181568A2/fr
Publication of WO2015181568A3 publication Critical patent/WO2015181568A3/fr

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01C—MEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
    • G01C21/00—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00
    • G01C21/26—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00 specially adapted for navigation in a road network
    • G01C21/34—Route searching; Route guidance
    • G01C21/3453—Special cost functions, i.e. other than distance or default speed limit of road segments
    • G01C21/3461—Preferred or disfavoured areas, e.g. dangerous zones, toll or emission zones, intersections, manoeuvre types, segments such as motorways, toll roads, ferries

Definitions

  • This invention relates to predicting attainable locations for a vehicle.
  • the invention may be thought of as determining the range of a vehicle .
  • the invention relates to predicting attainable locations and/or determining the range of an electric vehicle, perhaps over substantially all of the routes open to that vehicle from its location.
  • the invention may also relate to learning driver preferences on which range estimation can be based.
  • a method of determining a range for a vehicle using map data representing a map of a physical area comprising at least one of the following steps: i) obtaining location data determining the location of the vehicle within the map represented by the map data; ii) processing the map data to provide the energy requirement for substantially each possible route of the vehicle from its location to a destination within the area; and iii) processing the energy available to the vehicle to determine the range along each of the possible routes for the vehicle .
  • the steps i) to iii) may be performed in the order shown.
  • Embodiments providing such a method are believed advantageous because they allow a driver thereof to be able to determine whether or not any destination is within the vehicle's range without the need to enter a specific destination. That is, knowing the range allows a determination to be made as to whether a location is attainable or not. Such embodiments are therefore more convenient than the prior art and the technical problem of how to accurately estimate whether or not any possible destination is within the vehicle' s range is addressed.
  • embodiments may be arranged to automatically output attainable locations such that a driver may simply driver his/her vehicle and have the attainable locations updated as he/she drives.
  • possible route may be taken to mean a route ordinarily thought of as being passable by the vehicle that is using the method; ie the route is navigable. Accordingly, a car that is ordinarily driven along a road would not be ordinarily consider routes along farm tracks, unpaved roads and the like . A Heavy Goods Vehicle, or Passenger Service Vehicle, may not ordinarily consider routes along small single track roads and the like.
  • Embodiments providing the method of the first aspect may be thought of as summing the expected energy expenditure along each of the navigable segments within the map data.
  • the map data will contain a plurality of navigable segments with each segment representing at least a portion of a navigable road.
  • a segment of a road may exist between junctions at which a driver has a choice of direction.
  • a vehicle is on a segment its destination is known; i.e. the end of the segment, unless the vehicle were to stop on the segment itself.
  • substantially each is considered to mean that substantially each of the segments that are input to the method will be considered. Should segments present within the map data be suppressed from consideration by the method then the method will still consider substantially each of the segments input thereto.
  • at least some embodiments calculate the range along each of the possible routes within the map data, in real-time or at least substantially in real-time .
  • realtime is intended to mean as a driver drives the vehicle . It is likely to take a vehicle several seconds to reach the next junction of the navigable route along which it is travelling, which may be thought of as being a segment of the map data. It is convenient if the calculation of the range can be updated in the time that it typically takes the vehicle to reach the next junction (i.e. the time taken to traverse an average segment) and this may be thought of as being real-time.
  • real-time may mean on the order of substantially any of the following times: 2 seconds; 5 seconds; 10 seconds; 15 seconds; 30 seconds.
  • map data may constrain the map data to be considered by the method in order that the method can be performed in a short enough period.
  • short enough may be considered to mean in real-time as described above.
  • Constraining the map data may mean limiting the segments covered by the method.
  • the segments considered might be limited to those considered by a driver's driving preferences. For example, if the drivers driving preferences indicate that a driver does not drive on motorways then the method may not consider motorways.
  • constraining the map data may comprise limiting the map data considered to being a predetermined area.
  • the range of the vehicle may constrain the area considered.
  • area considered might also be a few tens of kilometres.
  • the map data covers a larger enough area to include the locations attainable by the vehicle.
  • the area may be a substantially circular area centred upon the vehicle. In such embodiments, the circular area may have a radius of substantially any of the following, in kilometres: 30, 40, 50, 60, 70, 80, 90, 100, 150, or the like.
  • the vehicle range may be plotted on a map providing a convenient user interface on which a driver can readily interpret the remaining range for the vehicle .
  • the map data comprises a plurality of navigable segments traversable by a vehicle and in which the method models the navigable segments within the map data as a network such that the trajectory through the network corresponds to a route for the vehicle.
  • the map data may be used to generate a model of the choices of a driver as a plurality of state-action pairs, with at least one pair being associated with a node of the network, and wherein the probability of a driver taking a particular trajectory is expressed as the probability of observing the corresponding state-action sequence through the network.
  • At least some embodiments may additionally use a learned characteristic to determine the probability of a driver taking a particular trajectory. Such embodiments are convenient as they allow the method to be tailored to a particular driver, which may be thought of as being a given driver.
  • the learned characteristic is updated from data which is recorded by a processing apparatus (or otherwise made available) as the vehicle, to which the method is being applied, is driven by a given driver.
  • Some embodiments may use a first Markov Decision Process (MDP) to model trajectory preferences and energy consumption.
  • MDP Markov Decision Process
  • a second MDP related to the first, and consisting of the same set of states, actions and rewards but with a transition model P' which encodes the transpose of the original road graph may also be used.
  • a vehicle including a processing apparatus, the processing apparatus being arranged to determine a range for the vehicle using map data representing a map of a physical area in which, in use, the vehicle is operated, the processing apparatus being further arranged to perform the following steps: i) obtain location data determining the location of the vehicle within the map represented by the map data; ii) process the map data to provide the energy requirement for substantially each possible route of the vehicle from its location; and iii) process the energy available to the vehicle to determine the range along each of the possible routes for the vehicle .
  • a machine readable medium containing instructions which when read by a processing apparatus cause that processing apparatus to perform the method of the first aspect of the invention.
  • a method of determining a range for a vehicle using map data representing a map of a physical area comprising: i) obtaining location data determining the location of the vehicle within the map represented by the map data; ii) processing the map data to determine the range of a vehicle to at least one destination, the range being based upon a learned characteristic providing a drivers driving preferences; iii) updating the learned characteristics as the vehicle performs the journey.
  • Embodiments providing such an approach account for driver-specific route preferences and/or route-specific factors by integrating over driver-specific distributions over possible trajectories to every possible destination.
  • Some embodiments may use a Markov Decision Process (MDP) to optimise an a priori unknown reward structure in order to update the learned characteristics as the vehicle is used.
  • MDP Markov Decision Process
  • the reward structure used in the MDP is a weighted linear combination of features.
  • an MDP framework allows for at least one of the following:
  • being able to calculate energy usage predictions for every destination for all routes means that the driver does not have to specify where he/she is going and what route would be preferred.
  • Such embodiments may therefore automatically display attainable locations as a driver uses his/her vehicle.
  • IRL Inverse Reinforcement Learning
  • IRL provides a principled, generalisable and efficient mathematical framework for updating the leaned characteristics.
  • other techniques could be used, such as regression (e.g. using Gaussian Processes) .
  • Some embodiments may additionally use the expected energy usage of the vehicle to travel along a route in order to determine the range / predict attainable locations.
  • the method determines the range / determines attainable locations along each of the routes open to a vehicle from its present location.
  • a confidence level with which a particular driver will be able to reach a particular destination is calculated.
  • the confidence level can also be referred to as a probability of attainability.
  • a current state of charge of the battery of an electric vehicle is used as an input for the calculation of the confidence level.
  • the confidence level is computed for all destinations in the map, specified by the map data, simultaneously. At least one of the following is considered in a model used for calculating the probability of attainability of a location:
  • the calculation of the confidence level allows the creation of a confidence level map (which may also be thought of as a probabilistic attainability map or a probabilistic range map) in which the confidence level is marked or otherwise recorded with reference to the map.
  • a confidence level map (which may also be thought of as a probabilistic attainability map or a probabilistic range map) in which the confidence level is marked or otherwise recorded with reference to the map.
  • the confidence level is calculated using a feature-based linear regression framework.
  • this allows for a computationally efficient implementation capable of providing real-time updates of the resulting confidence level map, wherein real-time may have the same meaning as described elsewhere herein.
  • one or more regions and/or lines may be marked or otherwise highlighted on the map to generate the confidence level map.
  • the regions and/or lines may be marked by the use of different colours, shadings, lines, line types or other visual indicators.
  • the regions relate to different ranges of confidence level that a destination can be reached. For example, a 90-95% confidence level may be presented, wherein locations within the region or regions are calculated as being attainable with a probability of 90-95%.
  • the regions are delineated by lines. In other embodiments, the regions may merge from one to the next.
  • the regions in such embodiments are marked by colour or shading variations.
  • a scale is provided indicating to which confidence level each visual indicator relates. The skilled person will understand that any percentage or percentage range may be chosen, and any number of lines and/or regions.
  • a vehicle including a processing apparatus, the processing apparatus being arranged to determine a range for the vehicle using map data representing a map of a physical area in which, in use, the vehicle is operated, the processing apparatus being further arranged to perform the following steps: i) obtaining location data determining the location of the vehicle within the map represented by the map data; ii) processing the map data to determine the range of a vehicle to at least one destination, the range being based upon a learned characteristic providing a drivers driving preference; iii) updating the learned characteristics with the vehicle performs a journey.
  • a machine readable medium containing instructions which when read by a processing apparatus cause that processing apparatus to perform the method of the fifth aspect of the invention.
  • the machine readable medium of any of the aspects and/or embodiments described herein may be any of the following: a DVD ROM/RAM (including -R/-RW and +R/+RW); a CD; a memory (including a USB flash drive; and SD card; a compact flash card; or the like); a hard drive (including both those with rotating platters and Solid State Drives); tape; an Internet download (including http; https; ftp; or the like); a wire .
  • a DVD ROM/RAM including -R/-RW and +R/+RW
  • a CD including a USB flash drive; and SD card; a compact flash card; or the like
  • a hard drive including both those with rotating platters and Solid State Drives
  • tape including http; https; ftp; or the like
  • an Internet download including http; https; ftp; or the like
  • Figure 1 shows three example range maps (a, b and c) generated by an embodiment
  • Figure 2 shows examples of likely trajectories between the same start and destination induced by three different route preferences
  • Figure 3a shows a trajectory spanning states and actions in an original road network (a);
  • Figure 3b shows a trajectory spanning states and actions for the trajectory of Figure 3a but in a transposed network
  • Figure 4 shows the forces acting on a car
  • Figure 5 shows velocity, acceleration and power demand profiles for three adjacent navigable segments respectively containing a speed limit, a stop sign and a slight elevation
  • Figure 6 shows how the mean predictive error changes as the number of training trajectories increases for a variety of models.
  • Figure 7 shows a flow chart outlining a first embodiment of the invention
  • Figure 8 shows a flow chart outlining a second embodiment of the invention
  • Figures 9A and 9B show examples of range maps generated by an embodiment in which probabilistic attainability is calculated.
  • Embodiments are based upon map data which represents a map showing an area through which a vehicle using an embodiment can move.
  • the map is considered to comprise a network of navigable paths composed of individual navigable segments (ie segments - s which are joined at intersections.
  • Each segment represents a section of a navigable path for a vehicle traversing the area covered by the map.
  • the segments will represent portions of a road, and for example may exist between adjacent intersections on a given road.
  • a segment might also represent a portion of a track, a path, rail, canal, tow-path, cycle path, or the like.
  • embodiments including the one being described, will limit the segments that are considered to those that are passable by the vehicle for which the method is being performed. As such, a method applied for a car will ordinarily consider road segments but would unlikely consider farm tracks, unpaved roads and the like.
  • map data representing the network of navigable paths together with associated route infrastructure can be readily obtained from, for example, community projects such as OpenStreetMap (www.openstreetmap.org).
  • OpenStreetMap www.openstreetmap.org
  • proprietary data such as obtainable from providers such as Tele-atlas, Navteq, or the like may be used to provide the map data.
  • Attainability may be determined by comparing an expectation of the amount of energy required to get to a place with the estimated state-of-charge of the battery of the electric vehicle. The expectation of the required amount of energy will depend upon range/distance, and may also depend upon environmental and driver preference factors, among other variables. In at least some embodiments, attainability may be thought of as a probability of reaching a specified place. Attainability may be calculated as an average over several routes, weighted by route preferences.
  • Attainability takes account of the differing energy requirements to travel the same distance, in different conditions and/or on different routes.
  • a processing apparatus is provided and arranged to process the various data that are described.
  • the processing apparatus is provided within the vehicle 100 such as seen at 400 in Figure 4.
  • the processing apparatus comprises a process such as an IntelTM , i5TM, i7TM, or the like; an AMDTM FX, AthlonTM; or the like.
  • the processing apparatus comprises a memory arranged to hold data and allow the processor to access, and process, that data in a manner that will be familiar to the skilled person.
  • the memory may comprise any suitable form of memory such as RAM (whether volatile or not), a hard-drive; a solid state drive, etc.
  • the memory may be local to the processor, accessible over a network by the processor, or a combination of local and networked. Indeed, it is conceivable that the memory is provided remote from the vehicle 100 and accessed via a wireless link (such as a GSM; UMTS; 3G; 4G; WIFI; or the like).
  • the vehicle 100 also comprises a location sensor arranged to determine the location of the vehicle (i.e. to obtain location data 700, 800) which is typically performed with reference to the map data.
  • the location sensor 402 comprises a receiver for a position networking and in particular the location sensor 402 is a GPS (Global Positioning System) receiver arranged to receive signal from the GPS system and determine the location of the vehicle 100.
  • the GPS receiver 402 generates location data which is passed to the processing apparatus 400.
  • a different location sensor may be provided.
  • a different positioning network may be used, such as any of the following: Global Navigation Satellite System (GLONASS); Galileo Positioning System; India Regional Navigational Satellite System; Chinese Compass Navigation System.
  • the location sensor 402 may be other than a positioning network and may for example be utilise cellular telephone information; gyroscopes within the vehicle 100; video camera systems arranged to determine the trajectory of the vehicle 100, or the like.
  • a vehicle 100 may comprise a plurality of the location sensors 402 listed here in order to take advantage of the relative merits of the different sensor technologies. Determining the range of a vehicle may be thought of as determining which destinations within the map are energetically attainable; that is which destinations are attainable with the energy stored within the vehicle. It is convenient to refer to an electric vehicle, such as an electric car.
  • the vehicle comprises further energy sensors arranged to determine energy data providing an estimate of the amount of energy remaining available to the vehicle.
  • the energy sensors may comprise at least the following: a current sensor, battery voltage sensor and/or temperature sensor. The energy remaining within the battery relates to the state of charge of the battery. Current state of charge may be calculated from the output of the energy sensors.
  • uncertainty in the estimation of the state of charge of the battery is accounted for by a noise term.
  • the origin of the noise or uncertainty is not important if the magnitude of the noise or uncertainty is known.
  • the vehicle may report 80%, and a +/- 5% uncertainty margin may be provided for use in the probability calculations.
  • different energy sensors may be used which may give an indication of the remaining volume of fuel (and therefore energy) on-board the vehicle 100.
  • Such energy sensors may for example monitor the volume of fuel (eg petrol (ie gas), diesel, hydrogen, LPG, or the like) remaining.
  • At least some of the embodiments are arranged to present a driver with an attainability map (which may be thought also as a range map) as shown in Figures 1(a), 1(b) and 1(c).
  • the vehicle 100 is shown at position Si on a map to which the underlying map data relates.
  • an envelope is displayed on the map to indicate how far the vehicle 100 can travel.
  • this envelope is shown as a shaded area on the map where segments within the shaded envelope are calculated to be reachable by the vehicle from the position Si.
  • other display formats might be used. For example segments that were reachable might be coloured or otherwise represented in a different manner to the segments that were not attainable.
  • Such learned behaviour is typically stored as a learned characteristic which is derived from data as a driver drives the vehicle to perform a journey.
  • the learned characteristic is stored within the memory of the processing apparatus and associated with an identified driver, which may be thought of as a given driver.
  • the vehicle 100 may have a plurality of driver profiles which specific drivers can select when using the vehicle 100. It will be appreciated that the learned characteristics are generated by the behaviour of a given driver and then used, in some embodiments, to generate a range prediction of the vehicle 100.
  • the learned characteristics may be derived from data including any one or more of the following: data providing the acceleration of the vehicle; data providing the velocity (i.e. speed) of the vehicle; data recording the selection of segments along which a driver chooses to drive the vehicle; the temperature; auxiliary systems (such as heating on/off, AC on/off, radio on/off, headlights on/off, or the like), road infrastructure including map information for individual segments (such as topology, elevation, number of lanes, existence of traffic lights, speed limit, or the like).
  • data from which the learned characteristics are generated may be thought of as giving a driver preference as to how the vehicle is driven.
  • the selection of segments along which a driver chooses to drive the vehicle may, in at least some embodiments, include an indication of a drivers preferred route selection. For example, it may indicate that a driver does not use freeways/motorways; prefers freeways/motorways, prefers roads with more bends, and the like.
  • energy consumption (voltage and current drawn) of the engine and the auxiliary systems are recorded. This enables computation of the actual energy demand for a particular segment.
  • the energy expended along a route can vary significantly according to the route chosen. Indeed, it is believed that the energy usage to move between two points (e.g. Sjand Sg) can vary by up to substantially 10% according to the route chosen. In the remainder of this description the following nomenclature is used: the destination (Sg) being attainable if the expected energy required to travel there E(s s , s g ) from the current location (s s ) for driver specific driving preferences ⁇ is less than the current state of charge of the battery, E soc .
  • embodiments aim to estimate the expected energy, E 9 (s s , Sg) to every destination, s g within the map.
  • E 9 expected energy
  • Sjg ( ⁇ ⁇ ⁇ )
  • each ⁇ ⁇ denotes a specific trajectory from s s to s g .
  • These trajectories can be substantially different and each one has associated with it a particular energy cost ⁇ ( ⁇ ).
  • driver's preferences over the set of trajectories as a condition of the likelihood of a particular trajectory being traversed, ⁇ ( ⁇ ⁇ ), on the set of driver-specific parameters.
  • Such a formulation computes, for every possible destination s goa i in the map, the expected energy consumed in getting there from the current location s s as
  • the embodiment being described presents a technique to mitigate these technical problems and uses a model where generalised route preferences of the driver are encoded as a particular policy in a Markov Decision Process (MDP).
  • MDP Markov Decision Process
  • the embodiment being described arranges this policy to optimise an a priori unknown, driver-specific reward structure, which can be learned and, over time, refined from trajectory data typically, using Inverse Reinforcement Learning (IRL).
  • INL Inverse Reinforcement Learning
  • MDP Markov Decision Process
  • the rewards express the energy cost of choosing a particular action in a particular state, E(si, a j ).
  • the transition model is deterministic such that a particular state-action pair leads to a particular next state with certainty, p(s j
  • Si, a k ) 1.
  • every trajectory has an associated total energy cost given by the sum of the individual costs
  • the embodiment assumes driver route preferences when driving to s g to result in a stochastic policy ⁇ (i.e. learned characteristic) a defining a probability distribution over possible actions p % (a,
  • the probability of the driver taking a particular trajectory can be expressed as the probability of observing the corresponding state-action sequence, such that:
  • At least some embodiments are arranged to use the map data to generate a model of the choices of a driver as a plurality of state-action pairs, with at least one pair being associated with a node of a network representing the navigable segments selectable by a driver, and wherein the probability of a driver taking a particular trajectory can be expressed as the probability of observing the corresponding state-action sequence through the network.
  • Equation 1 Computing E e (s s , S g ) as per Equation 1 is now equivalent to evaluating the value of policy ⁇ at state s s .
  • V 3 ⁇ 4 (s s ) can be found as a solution of a system of linear-equations as described in Sutton, R and Barto, A (1998, Reinforcement learning: An introduction, volume 1. Cambridge University Press). This teaching is hereby incorporated by reference and the skilled person is directed to read this teaching:
  • embodiments employing this method have an advantage in that they produce value V K (Si) and hence E e (si, s g ) for all starting states Si at once.
  • computation of a range map requires the opposite (E e (s s ,Si) for all destinations s ; an expected energy to reach every destination. It is hereinafter described how the embodiment being described computes E e (s s ,Si) at no extra processing cost.
  • s can be derived given a set of trajectories traversed.
  • This work is described further in Ziebart B.D.; Mass, A L; Bagnell, J A; Dey A. K.; 2008 "Maximum entropy inverse reinforcement learning; In Proceedings of the AAAI Conference on Artificial Intelligence; P1433-1438. This reference is hereby incorporated by reference and the skilled person is directed to read this paper.
  • the embodiment being described implicitly assumes a driver to optimise an a priori unknown cost function which leads to a particular trajectory to a given destination.
  • the embodiment being described employs a feature-based IRL formulation and expresses this cost as a driver-specific reward structure in an MDP defined over the road network. More specifically, this MDP is identical to the one described in the above apart from the reward structure, which is unknown and needs to be recovered.
  • the reward for a given state-action pair, Re(si, a j ) is assumed to be a weighted linear combination of features, f Si, a j , such that
  • ⁇ denotes the weight vector, which may be driver-specific and/or learned.
  • the features express various properties encountered when transitioning from one segment to another, such as the segment length, the time required to traverse it, the road class (e.g. motorway, dual carriage way, residential, etc.) and number of lanes, angle of turn as well as the number of full stops required due to, for example, the presence of traffic lights or stop signs.
  • the overall reward for a given trajectory, Re(Q, is computed as the sum of the rewards of the state-action pairs encountered along it,
  • This preference model gives rise to an equivalent stochastic policy in the MDP specifying the probability of taking action a j in state Si, which is proportional to the sum of probabilities of trajectories starting with the given action, such that
  • Maximum Entropy Inverse Reinforcement Learning can then be used to find the weight vector ⁇ for a given driver based on observed trajectories. This, as well as the computation of the driver specific policy can be carried out efficiently using a forward-backward algorithm, such as the one described in Ziebart B.D.; Mass, A L; Bagnell, J A; Dey A. K.; 2008 "Maximum entropy inverse reinforcement learning; In Proceedings of the AAAI Conference on Artificial Intelligence; P1433- 1438.
  • Embodiments using such a model are advantageous because they do not suffer from the label-bias problem as described in Lafferty, McCallum, and Pereira; 2001; "Conditional Random Fields: Probabilistic models for segmenting and labelling sequence data; In the Proceedings of the ICML; P336-343. As such, embodiments can therefore be computationally more efficient when evaluating this policy as discussed below.
  • the embodiment being described uses related MDPs consisting of the same set of states, actions and rewards but with a transition model P' which encodes the transpose of the original road network (step 702).
  • Figures 3(a) and 3(b) show a road network as it processed by the original MDP. It can be seen that the road network consists of 6 states: Si through s 6 . A path is shown on the network passing from state Si via s 2 via s 5 to s 6 .
  • Figure 3(b) shows the same road network but with the path therethrough reversed; i.e.
  • Equ. 8 implies that the probability distribution over trajectories T s g from s s to s g in the original MDPs is the same as the probability distribution over trajectories T' &s from s g to s s in the new MDPs. It follows, therefore, that E e (s s , Sg) in the original road network and ⁇ ' ⁇ (s g , s s ) in the transposed version are equivalent,
  • E e (s s , Si) can be computed for every Si at once by first constructing MDPs based on the transposed road network and then solving for the expected energy consumption ⁇ ' ⁇ (Si, s a ) from every starting state Si to the destination s s as described previously.
  • tint road network
  • the energy consumed over the course of a trajectory is modelled as the sum of energy costs of state-action pairs along the trajectory and given by Equ. 2.
  • This cost incorporates events specific to the transition, such as potentially stopping at an intersection or changing velocity according to the law of the land, as well as the energy required while traversing segment s t+ i itself.
  • To estimate these values we combine a canonical model of an expected velocity profile as well as a physical model of the resulting energy demand of the vehicle.
  • the embodiment being described models a driver's chosen velocity, v, and acceleration, v , as a segment is traversed by adapting the Intelligent Driver Model (IDM) described in Treiber M, Hennecke A and Helbing D; 2000; Congested Traffic States in Empirical Observations and Microscopic Simulations; Physical Review E62(2) P 1805-1824.
  • IDM Intelligent Driver Model
  • the IDM was originally applied in the context of a car-following scenario.
  • the embodiment being described employs the model by assuming the end of the segment to be a car moving at velocity v t equivalent to the desired velocity at the end of the segment. Given the road speed limit vo and distance to the end of the segment s, the acceleration of the driver is then given by the ordinary differential equation:
  • the energy cost of an action E(si, a j ) is obtained by integrating the vehicle's power demand P, during action a
  • the power demand of the vehicle is the sum of the engine power P eng and the auxiliary power demand P aux due to vehicle accessories such as lights, air conditioning, etc., such that
  • the engine power demand, P eng is a function of velocity and acceleration.
  • the embodiment being described assumes an efficient system such that
  • F eng is the force produced by the engine (F eng ) required to overcome forces acting on the vehicle at a given speed. As depicted in Fig. 4, these forces decompose into components due to acceleration (F acc) , friction (F fric ,i on) , air resistance (F air) and gravitation (F g ). ⁇ Constant Description Value
  • Table 1 Physical quantifies and valises ased to model energy consumption. 3 ⁇ 4 t «f ⁇ .,.- 4' i3 ⁇ 4is fe « ⁇ - ,;, :- ⁇ - F ⁇ >. ( 15) where
  • Fac t ion c rr ⁇ m ⁇ g is the rolling resistance
  • F g m ⁇ g- sin (a) is the hill-climbing force, where a is the angle of inclination of the hill.
  • Fig. 5 shows an example of the resulting acceleration, velocity and power demand profile of a vehicle traversing a typical segment of the road network.
  • the Figure shows how the velocity, acceleration and power are a function of the vehicle 500 position along three different segments: s t , s t+ i, s t +2-
  • s t the car 500 travels a constant speed (i.e. velocity) of above 30 km/h. It will be seen that after an initial period of acceleration 502 the velocity of the car is constant 504, there is zero acceleration 506 and consequently a zero power output from the engine.
  • the second segment s t+ i has a 30km/h speed limit in place and as such, it can be seen that as the vehicle 500 approaches the end of the first segment s t it decelerates to 30km/h and as such, there is a period of negative acceleration 508 before the car traverses the second segment s t+ i at a constant velocity 510. At the end of the second segment s t+ i there is a stop sign and the car 500 decelerates 512 to a standstill before accelerating into the third segment s t +2 which has a slight incline. After a short period the car 500 reaches a constant velocity but because of the increased F g force must now expend Power 514 to move up the hill.
  • the energy used by a vehicle for a particular trajectory or route is taken to be a random variable.
  • the distribution parameters are derived from real data collected from vehicles using the segments but this need not be the case.
  • Ei is therefore a mixture of distributions describing the energy consumed along possible trajectories, ⁇ ⁇ , with the mixing coefficients provided by the probability of a driver taking that particular trajectory
  • Equ. 16 an exact evaluation of Equ. 16 is often infeasible. Therefore, some embodiments re-approximate Ei with a normal distribution, which leads to computational efficiency and may be as specified in Sylvia Fruhwirth-Schnatter, "Finite mixture and Markov switching models", 2006.
  • the embodiment models the energy used be a vehicle to move along a trajectory as a sum over the traversed segments for that trajectory:
  • the parameters ⁇ 5 , a are modelled using a feature-based linear model similar to one employed to model segment rewards,
  • f s a again denotes a feature vector associated with every state-action pair and ⁇ ⁇ and ⁇ ⁇ denote driver specific weights parameterising the energy consumption.
  • the weights are learned by maximising the likelihood of observed energy consumption over the lifetime of the system by constraints- optimisation restricted by
  • Figure 6 shows the results of experiments that were conducted to test the performance of the embodiment being described, which use the energy model of Figure 5, and those experiments are now described in more detail.
  • range map generation crucially depends on the accuracy of the underlying estimates of energy consumption this discussion focuses on an evaluation on the relative error incurred in Ee(s sta it, s goa i)- Microsoft's GeoLife dataset is described in at least the following three documents: 1) Zheng, Y.; Li, Q.; Chen, Y.; Xie, X.; and Ma, W.
  • This GeoLife dataset contains GPS traces collected from a number of different users. Specifically, the experiments considered 50 different users, each having on average 100 trajectories spanning the urban area of Beijing, China. The road network and infrastructure information for this region cover about 100km 2 was obtained from the OpenStreetMap project. This road network is dense and contains a multitude of possible routes between any two places. The resulting MDP contains 80,000 states and 130,000 actions. As the dataset contains GPS traces only, these were segmented into individual trajectories based on time and position information.
  • Figure 6 shows the relative prediction error for 50 users as the amount of training data is increased . This is akin to more traj ectory information becoming available over time .
  • the Figure indicates that the prediction error decreases rapidly from that obtained using the initial model as the number of training traj ectories is increased. As information is added to the model beyond 16 training traj ectories improvements are more marginal until, at 28 training trajectories, the overall prediction error reaches ca. 6 : 5 %.
  • the formulation Ee(s sta r t , s goa i) is computed as an expectation over the entire set of possible traj ectories from start to goal state .
  • the benefit of this approach over considering, for example, only the most likely traj ectory given a driver' s preferences is also shown in the Figure .
  • the traj ectory from start to goal state corresponding to the maximum-likelihood estimate (MLE) is found by computing the shortest path through the road network using the learned segment cost and this is shown at 602 in the Figure . While the trend of the MLE solution is similar to that found when computing the full expectation in that the prediction error decreases as information is added to the system, the results suggest overall a performance decrease of ca. 1 % compared to using the entire distribution. In contrast, the approach of the embodiment described above computes an expectation over the entire distribution over traj ectories .
  • the Figure also shows two alternative models for path cost: travel time 604 and distance 606.
  • travel time 604 The Figure indicates that these commonly used estimates result in considerably higher overall prediction error confirming that drivers exhibit more complex preferences as to which route is taken.
  • Figures 9A and 9B show examples of confidence level maps 900, 920 in which information has been added to the map which highlights the probability (ie a level of confidence) with which a driver can reach a destination.
  • the skilled person will appreciate that embodiments are making a prediction about the range of a vehicle and as such, it might be advantageous to provide the level of confidence in that prediction.
  • the position of the vehicle is marked by an arrow 910.
  • the marked regions on the map indicate the attainable range with 90% (902), 50% (904) and 10% (906) probability (ie a 90% confidence; a 50% confidence; and a 10% confidence - where the lower the number the less likely it is that the destination can be reached).
  • the region of the map outside of these marked regions (902, 904, 906) is therefore attainable with a probability of less than 10%.
  • confidences other than 90%, 50% and 10% may be shown.
  • Figure 9B demonstrates that there may be more than one distinct region corresponding to a certain probability of attainability.
  • Region 904 is split into 904a and 904b.
  • the presence of a region such as 904a may indicate the presence of a hill or other aspect of road infrastructure which increases the energy requirement for reaching a particular location or set of locations.
  • a particular region 904a may have a lower probability of attainability than some locations at a greater distance from the vehicle position 910.
  • This principle also applies to the example range maps shown in Figure 1 , and to maps generated in at least some other embodiments.
  • probabilistic range maps require additional calculations and inputs.
  • Such embodiments may use noise terms, also known as uncertainty or error terms, in estimation of the remaining energy in the battery and of the energy requirement for reaching a destination are used to provide a confidence level.
  • Such embodiments may therefore use the energy model in which the energy usage of the vehicle is modelled as a distribution may therefore by used and as described above with reference to equations 16 to 25.
  • the probability that a particular destination can be reached is calculated for every destination in the map.
  • a probability pi that required energy Ei required to travel to destination Si is at most the available energy E soc is estimated .
  • destination is considered attainable if this probability is at least p t
  • Figures 9A and 9B demonstrate attainability of destinations for different levels of p t . Equ. 30 can then be expressed in terms of the cumulative distribution function of the standard normal distribution such that
  • these parameters are also learnt as part of the model .
  • Embodiments compute parameters ⁇ and Oi to allow for the generation of probabilistic attainability maps as shown in Figure 9.
  • G s a factorising over state-action (s,a) space .
  • the generation of an attainability map computes an expectation of G s a over the distribution of all possible traj ectories ⁇ from a starting position to a particular destination
  • Polynomial-time algorithm Algorithm 1 indicated by Q( » ), given a reward structure R, computes this expectation from a starting position s sta rt to every possible destination in the map, such that

Landscapes

  • Engineering & Computer Science (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Remote Sensing (AREA)
  • Automation & Control Theory (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Navigation (AREA)

Abstract

La présente invention concerne un procédé de prévision des emplacements atteignables pour un véhicule, sur la base de son emplacement actuel, la prévision utilisant des données cartographiques représentant une carte d'une zone physique autour de l'emplacement actuel et les données cartographiques comprenant une pluralité de segments navigables, qui peuvent être traversés par un véhicule, le procédé comprenant les étapes consistant à obtenir des données de localisation déterminant l'emplacement du véhicule à l'intérieur de la carte représentée par les données cartographiques ; estimer l'énergie requise pour traverser chacun des segments navigables au sein des données cartographiques ; traiter les données cartographiques conjointement avec l'estimation d'énergie pour fournir l'énergie nécessaire pour sensiblement chaque trajet du véhicule de son emplacement actuel à chaque destination à l'intérieur de la zone ; surveiller une source d'énergie du véhicule pour déterminer l'énergie restant dans celui-ci ; traiter l'énergie restante disponible avec les besoins d'énergie pour sensiblement chaque trajet possible pour déterminer des emplacements atteignables pour le véhicule ; et émettre les emplacements atteignables pour le véhicule.
PCT/GB2015/051575 2014-05-30 2015-05-29 Prévision d'emplacements atteignables WO2015181568A2 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
GBGB1409620.0A GB201409620D0 (en) 2014-05-30 2014-05-30 Range determination
GB1409620.0 2014-05-30

Publications (2)

Publication Number Publication Date
WO2015181568A2 true WO2015181568A2 (fr) 2015-12-03
WO2015181568A3 WO2015181568A3 (fr) 2016-03-10

Family

ID=51214470

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/GB2015/051575 WO2015181568A2 (fr) 2014-05-30 2015-05-29 Prévision d'emplacements atteignables

Country Status (2)

Country Link
GB (1) GB201409620D0 (fr)
WO (1) WO2015181568A2 (fr)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111007858A (zh) * 2019-12-23 2020-04-14 北京三快在线科技有限公司 车辆行驶决策模型的训练方法、行驶决策确定方法及装置
US10902347B2 (en) 2017-04-11 2021-01-26 International Business Machines Corporation Rule creation using MDP and inverse reinforcement learning
CN115134748A (zh) * 2022-02-17 2022-09-30 恒安嘉新（北京）科技股份公司 确定漫入漫出用户的方法、装置、电子设备及存储介质

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114459498A (zh) * 2022-03-14 2022-05-10 南京理工大学 基于强化学习的新能源车充电站选择及自适应导航方法

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8874367B2 (en) * 2011-10-14 2014-10-28 Equilateral Technologies, Inc. Method for estimating and displaying range of a vehicle
CN104583722B (zh) * 2012-06-29 2018-06-08 通腾发展德国公司 用于路线搜索的设备及方法

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10902347B2 (en) 2017-04-11 2021-01-26 International Business Machines Corporation Rule creation using MDP and inverse reinforcement learning
US11003998B2 (en) 2017-04-11 2021-05-11 International Business Machines Corporation Rule creation using MDP and inverse reinforcement learning
CN111007858A (zh) * 2019-12-23 2020-04-14 北京三快在线科技有限公司 车辆行驶决策模型的训练方法、行驶决策确定方法及装置
CN115134748A (zh) * 2022-02-17 2022-09-30 恒安嘉新（北京）科技股份公司 确定漫入漫出用户的方法、装置、电子设备及存储介质

Also Published As

Publication number Publication date
GB201409620D0 (en) 2014-07-16
WO2015181568A3 (fr) 2016-03-10

Similar Documents

Publication Publication Date Title
US11650066B2 (en) Systems and methods for variable energy routing and tracking
US10663305B2 (en) Map matched aggregation for K-anonymity in trajectory data
US11562168B2 (en) Clustering for K-anonymity in location trajectory data
Thibault et al. A unified approach for electric vehicles range maximization via eco-routing, eco-driving, and energy consumption prediction
Yuan et al. Driving with knowledge from the physical world
US8554473B2 (en) Energy efficient routing using an impedance factor
EP2428770B1 (fr) Système de navigation de véhicule
US9891062B2 (en) System for estimating a cruising range of a vehicle
US8538686B2 (en) Transport-dependent prediction of destinations
Krumm et al. Predestination: Where do you want to go today?
US20200019585A1 (en) Triangulation for k-anonymity in location trajectory data
EP3561453B1 (fr) Procédé, appareil et produit programme informatique permettant de déterminer la probabilité d'un itinéraire
Ondruska et al. The route not taken: Driver-centric estimation of electric vehicle range
CN103542858A (zh) 车辆到达目标能力评估方法，数据库生成方法、导航系统
US11802775B2 (en) Methods and systems for determining safe return range
JP6462264B2 (ja) 目的地予測装置、目的地予測方法、及びプログラム
US9983016B2 (en) Predicting short term travel behavior with unknown destination
US20180045527A1 (en) Systems and Methods for Predicting Vehicle Fuel Consumption
WO2015181568A2 (fr) Prévision d'emplacements atteignables
US20200141738A1 (en) Navigation system, apparatus and method for associating a probe point with a road segment
Gebhard et al. Range prediction for electric bicycles
US20200386569A1 (en) Trajectory sampling using spatial familiarity
CN116194935B (zh) 用于确定地理区域中的车辆的导航简档的方法和设备
CN112183871A (zh) 基于空气指数的城市交通诱导系统
Suraji et al. Smart Route Choice Based on Google Maps Application in Urban Road Network

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 15726273

Country of ref document: EP

Kind code of ref document: A2

NENP Non-entry into the national phase in:

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 15726273

Country of ref document: EP

Kind code of ref document: A2