GB2534020B - A device to estimate the geolocation of a moving signal emitter/receiver - Google Patents

Description

A device to estimate the geolocation of a moving signal emitter/receiver

The present invention is an improved method of geolocation using Received Signal Strength (RSS).

Geolocation is an important requirement for both users and operators of mobile communication networks. Typical motivations driving the requirement for geolocation include compliance with the E-911 emergency service in the US, and providing location based services to users (such as finding the location of the nearest restaurant or coffee shop, or providing location aware weather or traffic). A number of techniques exist that attempt to solve the geolocation problem, including GPS, Direction Finding (DF), Time of Arrival (TOA)/Time Difference of Arrival (TDOA)/Angle of Arrival (AOA)/Time Sum of Arrival. However, methods of geolocation that are reliant on timing information, such as TDOA/TOA require costly timing hardware or complex antenna arrays in the case of AOA. The hardware necessary for the use of techniques such as DF also requires a great deal of careful calibration. Some of these common geolocation techniques also rely on knowledge of the structure of the signal of interest for correlation.

There also exist methods to determine the location of a radio frequency (RF) emitter using RSS values. A received signal strength indicator is a simple measurement of the amount of RF energy received at a receiver from a signal broadcast by an emitter. The principle advantage of RSS over other techniques is the low level of cost and complexity of the sensors; unlike methods that are reliant on timing information, sensors measuring an RSS value do not require any costly timing hardware (as is the case with TDOA/TOA) or complex antenna arrays in the case of AOA, which reduces the cost of manufacture significantly. It also avoids the careful calibration required for DF. Geolocation using RSS means there is no need to rely on knowledge of the structure of the signal of interest for timing correlation, unlike other common non-RSS based geolocation techniques.

An important factor in the design of geolocation systems is whether the system is to be used with cooperative or non-cooperative emitters. A typical RSS based geolocation system consists of one or more sensors or base stations distributed around the area of interest, taking measurements of transmissions from the emitter to be located. In some cases (such as GSM®) the roles are reversed, with the mobile transceiver now functioning as a receiver and taking measurements from a number of "beacons" that transmit a constant signal. In the case of a cooperative emitter, the base stations have knowledge of the transmission schedule, can make reasonable assumptions about the antenna type and polarization, and crucially have knowledge of the emitted power level. As well as being inaccurate, most RSS-based geolocation systems are unable to geolocate non-cooperative emitters.

Existing methods to determine the location of an RF emitter using RSS values can be broadly divided into those relying on geometric models, and fingerprint based models relying either on a database of physical measurements or on predicted readings generated through use of site-specific modelling tools such as ray tracers. The methods based on geometric methods and non-site specific model suffer from the principle disadvantage that they do not incorporate fading effects present due to terrain features or buildings; this means that the predictions offered by the models, and thus the geolocation results, are often very inaccurate; for example, there is no way of telling whether a low R.SS value is caused by the emitter being geographically far away from the receiver, or if it is due to a building in between the emitter and receiver, as illustrated in Figure 1.

Database based methods also suffer from similar issues. A large factor in the performance of a database based geolocation system is the quality of the underlying database. The two principle ways of creating a database to be used for geolocation are to either conduct an extensive measurement campaign (essentially, measure the received signal strength at many points over the area of interest), or to use a site-specific propagation model to generate predicted received signal strengths at many points over the area of interest. Once the database has been gathered or generated, database-based geolocation algorithms select the database point that most closely resembles the observed received signal strength vector as the predicted most likely location of the emitter, according to a similarity metric. However, there are often a number of database points that match the observed readings to a similar degree of accuracy, due to imperfections in the database, and shadowing effects. These points are often not geographically close giving rise to a number of potential locations for the emitter. The variety of potential geolocation results using traditional RSS-based techniques is shown in Figure 13. As such, using RSS is not currently a reliable way to accurately geolocate an emitter.

The question then becomes one of how to resolve these ambiguous geolocation results into a true location of the emitter. This is the problem that the invention presented here solves.

Accordingly there is provided a means to estimate the geolocation of a moving signal emitter/receiver comprising:- a) A plurality of receivers or transmitters for receiving or transmitting signal data to or from the moving signal emitter/receiver; b) Collecting the signal data at the signal receivers or at the moving signal emitter/receiver at multiple points in time and generating a timestamped Received Signal Strength (RSS) vector for each of these points in time; c) means to determine the impact of features in an area upon the RSS, generating predicted RSS over said area, and using this to generate predicted Received Signal Strength (RSS) vectors at more than one location within said area; d) Estimating probable positions over time for the moving signal emitter / receiver based upon steps b) and c);

Characterised in that e) there is provided means to generate a series of candidate routes denoting the moving signal emitter / receiver movement; f) using an Emitter Movement Model (EMM) to estimate the probability of any of the series of candidate routes being taken; g) means to estimate the correct emitter movement model to apply to the moving signal emitter/receiver; h) means to compare each timestamped Received Signal Strength (RSS) vector, each of which corresponds to a location within the area, with the predicted Received Signal Strength (RSS) vector corresponding to the same location within the area, to provide a similarity metric; and i) means to find the most probable of all of the candidate routes to determine the most likely route taken, by finding the product of the probability of the route being taken (in f) with the summation of the similarity metrics (in h).

Performing geolocation using RSS according to the invention is applicable even when there is no knowledge of the modulation scheme used, and involves less computationally expensive signal processing routines.

The present invention discloses a method of geolocation which is accurate and reliable by solving problems with existing RSS-based approaches, yet simpler and cheaper than non-RSS approaches.

Another of the advantages of the invention is procedure is its applicability to geolocation of non-cooperative emitters; the received signal strength difference embodiment of this technique incorporates the assumption that we do not possess concrete knowledge of the transmission power of the device.

The principle assumption is that we have some knowledge of the movement capabilities of the emitter; e.g if it is in a car, carried on foot, or some information on how it is moving, such as if it is moving from one destination to another as quickly as possible, or if it is on a route covering multiple destinations. This enables us to build an Emitter Movement Model (EMM), which may be as simple or sophisticated as desired. By combining knowledge of how the emitter moves with a set of potentially inaccurate geolocation results gathered over time, a more accurate idea of the route followed by the emitter and its location can be inferred. The nature of perceived movement may also be used to refine the EMM selected - for example if all the candidate routes involve large distances between timestamps it is likely that the emitter is on or in a vehicle; thereby allowing the choice of EMM to be refined.

The most basic implementation of an EMM may be one which tests if a candidate route is possible (for example by testing the route against the known max speed of the emitter), returning a simple binary answer to discount impossible routes. More complex EMMs could be defined, which instead of a binary result return a probability estimate against one or more factors. These more complex EMMs may consider a multitude of factors such as road networks, preferences for certain routes or roads, tendency to stop or change speed, tendency to change direction, traffic patterns, population statistics (e.g. movement from one densely populated areas to another may be more likely than movement to or from more sparsely populated areas), weather (may affect choice of transport or route), time of day (may affect route), or essentially any other information which may impact upon the route taken.

There is no limit on the number of factors or complexity of factors which could be included in the EMM.

The EMM returns a likelihood estimate for a given sequence of movements, to further inform the geolocation algorithm, and the formulation of the geolocation problem as an multi-objective optimisation task, solvable through application of a broad range of existing optimisation techniques. The likelihood estimate may be a simple binary result (yes/no or 1/0) in the case of a simple EMM or may be a more sophisticated probability estimate.

The invention will now be described with reference to the following drawings.

Figure 1 - Shows the inaccuracy of known RSS values due to either geographical distance of the emitter (1) from the receiver (as in the case of receiver 101) or due to signal attenuation because of an intervening building (20) (as in the case of receiver 102). This attenuation problem caused by physical and man-made geography makes RSS an inaccurate method of geolocation in all but very benign environments.

Figure 2 shows a potential system that could be used to provide data for a geolocation system as according to the invention. A number of Receivers (101-105) are deployed in an area, represented by triangles. Each of these receivers is equipped with a radio receiver, and observes broadcasts from the emitter. The "area of interest" is labelled as the box (30). The mobile emitter in this case is represented by a star shape (1), with the path taken represented by a trail of dots behind it (3). Receivers (101-105) make observations on the emitter (1), and transmit measurements (4) back to the processing centre (8) together with a timestamp denoting when they were observed. Whilst this is one possible embodiment of a system that could be used to provide input to the system, it is by no means the only method through which received signal strengths measurements could be obtained, and is provided merely as an illustrative example of one possible system.

Figure 3 shows an alternative deployment of the system where the receiver is based on the mobile device to be geolocated (3) following a path (3), taking received signal strength measurements on multiple reference emitters (201 - 205) and sending the results (4) to a processor (8).

Figure 4 shows an example embodiment of the system, how they are related and the data flows between them. The algorithm (18) is presented with a sequence of observed received signal strengths (16) gathered from receivers located at the locations (12). These could be presented to the algorithm as a list of vectors containing received signal strengths measured in dBm (such a table is shown in Figure 5.) The propagation model (17) is used to calculate predicted received signal strength vectors at each of the possible states in (11). The propagation model (17) may be a simple model such as the free-space loss model, a more advanced site-specific model such as those based on ray tracing principles, or one based on empirical measurements such as the COST, Okumura or Hata models. Other inputs to the example embodiment include the Emitter Movement Model (14), an optional terrain map (13), a similarity metric (15), and a sequence of observed measurements (16). Once the system has been presented with the required inputs, the system solves the optimisation problem, presenting the sequence of states 5 as the output of the algorithm (19).

Figure 5 shows an example set of received signal strength measurements, in a form that could be presented to the algorithm. Each RSS vector is numbered, and has an associated timestamp. The measurements for each receiver are then presented in units of dBm.

Figure 6 shows an example set of received signal strength difference measurements. In some scenarios it may be desirable to operate on differential received signal strengths (that is, the differences between signal strengths observed at various receivers). For example, the difference in magnitude of the signal observed at Receiver 1 and Receiver 2 is listed in the column labelled Rssl-Rss2.

Figure 7 shows a typical propagation environment showing shadowing due to buildings (20, 21, 22), a problem which makes geolocation using RSS difficult. A signal (30) is reflected (31) from a building (21). The indirect propagation path (30,31) will affect the RSS signal received at the receiver (201) from the emitter (1)

Figure 8 shows a typical propagation I radio map.

Figure 9 shows graphically how genes may be combined if a genetic algorithm is used to solve the optimisation problem.

Figures 10, 11, 12 and 13 show sample results from a prototype system (approximate width of area shown in figures 10, 11 and 12 is 300m).

Figures 10a, 11a, 12a and 13a are the actual outputs from the prototype system, from which figures 10, 11, 12 and 13 are drawn respectively. The features highlighted on figures 10, 11, 12 and 13 accord with the corresponding features in Figures 10a, 11a, 12a and 13a respectively.

Figures 10, 11 and 12 show the localised geographical area in which the emitter travelled (50) which fully or partially contains a number of buildings (60, 61, 62 and 63 in figures 10, 11 and 12).

Figure 10 shows the actual emitter path 71 travelled (as it happens, following a road within the localised area 50.)

Figure 11 shows the predicted emitter path 72, which falls within the same localised area 50.

Figure 12 shows the actual emitter path 71 and the predicted emitter path 72 overlaid on the same drawing.

Figure 13 shows the wider geographical area in which the emitter was being geolocated in the prototype system. The ambiguous results (potential geolocations of the emitter) from a traditional RSS-based geolocation system are shown as intersecting arrows 80, some of which are indicated. These ambiguous results 80 are spread across the wider geographical area despite the emitter movement having occurred within a localised area 50 as previously referred to in Figures 10, 11 and 12. Only one of the ambiguous results 80 falls within the localised area in which the emitter travelled 50 and is indicated as 81.

The requisite components for operation of a geolocation system according to the invention are described below.

Emitter Movement Model

With reference to Figure 4, an Emitter Movement Model (EMM) (14) computes the likelihood of a certain sequence of state transitions occurring, state transitions representing the emitter moving from one cell in the map to another. This model may take the form of a simple probability distribution of speed and direction deltas, potentially derived from a recorded GPS trace of an example emitter. It may also be combined with an annotated map (13), indicating the position of roads, junctions, and buildings; these may also be used to provide additional information about the likely behaviour of the emitter. For example, the emitter movement model may rate the probability of an emitter moving at high speed along a road as being more likely than in a building, and may discount the possibility of an emitter moving directly through an obstacle. This probability may be in the form of a binary result (yes/no or 1/0) or a infinitely variable probability. In any case, it takes the form of a model that is presented with a sequence of state transitions representing a path taken by an emitter over the map, and yields a likelihood for the path.

Deployed Signal Receivers / Emitters

This phase deals with the deployment of multiple sensing/receiving nodes over the area of interest, and generation of radio maps (Figure 8) that will enable geolocation using the propagation model. Transmitters/Receivers are placed at one or more locations that cover the area of interest, or arrangements are made for existing sensors to deliver time-stamped received signal strength measurements.

Depending on the configuration chosen, a plurality of Signal Receivers (Figure 2, Receivers 101-105) or reference transmitters (Figure 3, Transmitters 201-204) are distributed to cover the area of interest. Alternatively, arrangements are made for existing signal receivers or signal transmitters to deliver time-stamped received signal strength measurements. For ease of understanding, throughout this description we will assume that we are geolocating a mobile emitter where a number of base stations are acting as receivers (Figure 2); it will be appreciated by anyone skilled in the art how the alternative configuration could be used (Figure 3) where the moving object is acting as a receiver and the base stations are acting as emitters.

Observed signal A radio signal broadcast from a wireless emitter (Figure 2, 1) is observed at two or more wireless receiving devices (Figure 2, 101-105) over a period of time, or alternatively signals broadcast from multiple reference emitters (Figure 3, 201-205) are observed at the mobile device (figure 3, 2). At an interval t, the received signal strength at each receiver is sampled. The observed signal strength at each receiver are combined into a received signal strength vector (Figure 5) for each time interval and sent to the processor (Figures 2 & 3, 8)

Propagation Model

The preferred propagation model is one that takes into account site specific shadowing and fading parameters, for example a model based around the principles of ray tracing or similar propagation prediction mechanisms. The advantages gained by using a ray tracer over a more simplistic model (e.g. Free space/Okumura-Hata) are that the effects of shadowing are more closely modelled, resulting in a closer match between observed signals and the output of the model.

Figure 7 contains an example of an urban propagation environment. The direct path between the emitter and receiver is obstructed by 20. If a simplistic, non-site-specific model were used (without knowledge of the path between emitter and receiver), it would over predict the received signal strength at the receiver, as it would not account for the losses caused by the radio waves passing through the concrete or brick of the building. A ray tracer is capable of solving this problem; the ray tracing engine has knowledge of the surrounding environment, including the location of buildings, trees, and variations in the underlying terrain. A common method of operation for ray tracers is to "launch" rays in all directions at regular intervals from the emitter; the interactions that these rays have with the environment is then modelled, and predictions about the RSSI at a given location can be made with a higher degree of confidence than non-site-specific models. Figure 7 shows an example of one such ray; the ray (30) is launched from the emitter towards building (21), where it reflects (31) or is scattered by the surface of the building. A secondary ray then arrives at the emitter, taking into account the losses from the reflection. The power observed at the receiver may be the product of many interacting rays in more complex propagation environments (e.g. cities). This is why geolocation using RSS is difficult to do with any accuracy outside of a benign environment.

Algorithm Description

At a high level, the following equation represents the output of the proposed system: S' ( V ,· ' Ά ' ... i=i

Equation 1 - High Level Description • M represents the set of all possible locations on the map; this may correspond to points or cells on a grid, or arbitrarily shaped areas. • S represents a sequence of states denoting the emitter moving from one point or area on the map to another, drawn from M. • $ is the set of states that is the output of the process, representing the most likely route taken by the emitter. • model(S) is a function that takes a sequence of states S (or "route"), and returns a likelihood estimate based on some model of emitter movement. • N is the number of observations made on the RSS of the emitter, each observation consisting of one measurement per receiver. • obsi jS the set of observed signals

• is the location or area on the map M of the state in the candidate path S . the predicted received signal strength vector at the location according to a propagation model or received signal strength database • sim is a similarity metric that computes the similarity between an observed set of received signal strengths, and a predicted set of value • l/l/p is a weighting factor applied to the propagation model similarity results • Wm is a weighting factor applied to the movement model results.

Similarity Metrics

In practise, any function that takes two vectors and returns a distance between them could be used. There are many different metrics that may be used to determine the similarity between predicted and observed received signal strength vectors.

For the purposes of illustrating the concepts behind the invention, two such metrics are outlined below.

The Euclidean distance may be used as a similarity metric as follows:

Equation 2 - Euclidean Distance Metric

Where • There are K receivers . o&sujS the rss observed at receiver r at the state on the route

• £'ri?£^r is the predicted RSS value of receiver r according to a radio model or database at the location of the state on the route.

Or alternatively, the cosine similarity metric may be used:

Equation 3 - Cosine Similarity Metric

Where the term definitions remain the same. Note that there are infinitely many similarity metrics that could be defined, as any metric that provides a measure of similarity between two vectors could conceivably be used.

Example movement model

There are many ways to calculate the likelihood of an emitter following a particular path. One of the simplest movement models is the model that discards impossible routes.

Equation 4 - Simple movement model

Where • S is the list of states representing the proposed route taken by the emitter ► u-Jjs the physical distance between states (e.g. Euclidean • sMs the timestamp corresponding to the observed state • maxSpeed is a predetermined maximum speed that the emitter is capable of

Used in this form, the model returns zero whenever a route incorporates a manoeuvre that violates the maximum speed of the emitter; in Equation 1, this results in the term for the given state sequence resolving to zero. More complex models of emitter movement are also possible.

Extended Movement Model A more complex model of emitter movement might utilise data from modes of transportation that move in a similar way to the mode of transportation that the emitter that will be located by the system is mounted on. The basic principle behind this is to take a movement trace recorded from a vehicle or other mode of transport that we believe to behave in a similar manner to the vehicle or other mode of transport, and from this generate summary statistics that are used to assess the similarity to predicted paths generated by the algorithm. This movement trace could be obtained by recording a time-stamped series of GPS locations using (for example) a handheld GPS unit or smartphone with appropriate software installed, attached to the example emitter. Statistics on these GPS stamped locations are then computed, and then compared to the statistics of the paths generated in the geolocation process.

It is possible to calculate various statistics about a GPS trace path; for example: • Average (mean)/standard deviation of speed over journey segment. A high average speed may correspond to movement in a vehicle, a low mean speed might correspond to the emitter being carried by a pedestrian. The standard deviation of the speed captures information about whether the vehicle was travelling at a constant speed, or at a number of different speeds. • Average (mean) speed change between samples. A high average speed change between samples would be indicative of rapid acceleration/deceleration; a low average speed change indicates that the emitter was travelling at a more constant speed. • Average (mean)/standard deviation of direction change between samples. This provides a measure of how often the emitter changes direction. • Kurtosis values for the heading and speed changes between samples. Kurtosis gives an indication of how much of the variability in the sample is due to a small number of larger variances, or a large number of smaller differences from the mean.

Once the statistics from the example movement model have been generated, they can be stored in a table or database. The same statistics can be calculated on the paths generated as part of the geolocation process, and compared to those in the database. Generated paths can then be scored according to the similarity of their summary statistics to the entries in the database. An example implementation of this technique is described below.

Example Propagation Model

Whilst it is anticipated that the invention will be used with some form of database, or a site-specific propagation model such as a ray tracer, it is possible to demonstrate the concept with a simple propagation model, as follows. R(dS?n)= Pe(dSm>- lteioiff — 1

Equation 5 - Free-space loss propagation model

An alternative to this is the two-ray model that incorporates information about the transmitter and receiving nodes, as described later ftjftj = ptGt£r-^-

Equation 6 - Two-Ray Model

Where « rJs the power at: the receiving antenna • pe is the transmitted power « >%. is the transmitter gain • is the receiver gain • ftf is the height of the transmit antenna • is the height of the receive antenna • d Is the transmit/receive separation distance

Example Embodiments

An example embodiment of the invention is shown below, utilising the Euclidean distance metric and free-space propagation model, and assuming a state space consisting of points defined in a 2D Cartesian coordinate system.

Equation 7 - Implementation with Euclidean Distance, simple movement model and Free-space Loss

Or alternatively, using the two-ray model incorporating antenna gains and heights:

Equation 8 - Example with Euclidean Distance, Simple Movement Model, and Two-Ray propagation model

Where the following definitions apply:

• H is the map containing all possible states/locatlons • N is the number of observations made * K is the number of receivers ♦ S is a sequence of H states drawn from H representing a route followed by the emitter • 5 is the sequence of states that correspond to the most likely route taken by the emitter * s£ is the x coordinate of the ii?1 state in S (e.g. units of metres) ♦ is the y coordinate of the state in S (e.g. units of metres) * sJs the time at which the A observation in S was made (e.g. units of seconds) « moxSpeed is the maximum speed of the emitter (units of metres per second) * ofojs the received signal strength observed at the/ril receiver at the x j " “ • ί^ observation, measured in units of dBm ♦ dsls the reference distance at which the signal strength measurement pswas made on the emitter (e.g. units of metres) ♦ As(ds?ra)is the signal strength Pe converted to dBm • the path loss exponent for receiver / to the location of state s • r, is the x coordinate of the /ίλ receiver • r, is the y coordinate of the Λ* receiver • PJs the transmitted power ♦ and are the respective gains of the transmitting and recei ving antennas * fej. and fy, are the respective heights of the transmitting and receiving antennas

Alternatively, in the case where the receiver is co-located with the mobile device to be located as in Figure 3, the problem could be formulated as follows:

Equation 9 - Implementation with Euclidean Distance, simple movement model and Free-space Loss

Where is the emitted power level of the reference transmitter measured in dBm at distance^, and all other term definitions are identical to those in Equation 7.

Example extension using differential received signal strength A further example embodiment of the invention is shown below, utilising the Euclidean distance metric and free-space propagation model, and assuming a state space consisting of points defined in a 2D Cartesian coordinate system. This model is an extension of that detailed in, utilising differential received signal strength instead of absolute. Received signal strength difference based models offer the advantage that the transmit power is not necessarily needed; it is possible to insert an arbitrary value in the equation for

Equation 10 - Implementation with Received Signal Strength Difference, Euclidean Distance, simple movement model and Free-space Loss

Where the following definitions apply:

* M is the map containing ail possible states/locahons * N is the number of observations made * K Is the number of receivers » S is a sequence of N states drawn from M representing a route followed by the emitter * sis the sequence of states that correspond to the most likely route taken by the emitter * s, is the x coordinate of the irii state in S (e.g. units of metres) * is the y coordinate of the ii& state in S (e.g, units of metres) * sjs the time at which the d* observation in S was made (e.g, units of seconds) * mast-5??eeiSs the maximum speed of the emitter (e.g. units of metres per second) * o&^ .is the received signal strength observed at the ji,!t receiver at the ΐίΑ observation, measured in units of dBm. * 4jis- the reference distance at which the signal strength measurement p8was made on the emitter (e.g. units of metres) * pCi(dSm-)is the signal strength pc converted to dBm * is the path loss exponent for Receiver / to the location of state s » η is the x coordinate of the jiis receiver * jsr * n is the y coordinate of the ?ίλ receiver

Solving the optimisation problem

Clearly, for any non-trivially sized state space or path length, discovering the optimal solution is not trivial. There are many potential methods for finding a near optimal solution, including genetic algorithms and combinations of depth and breadth first search. In this section, an example implementation utilising a genetic algorithm is presented.

Genetic Algorithm

The principles behind genetic algorithms used for optimisation are well known and understood in the field, and can be found in any number of textbooks on machine learning. The principle idea behind the genetic algorithm is to start with a large number of randomly generated potential solutions to the problem (in this case, a solution would consist of a randomly generated path across the area of interest, representing the predicted path of the emitter). The randomly generated solutions are then scored according to a fitness function, ranking the solutions by a) their similarity between predicted and observed RSS values and b) The path geometry scored by the emitter movement model. The top n solutions are then combined in various ways to produce a new set of results, which are then passed forward into subsequent generations where the process is repeated. There are three main steps to implementing a system based around a genetic algorithm; 1. Formulate an appropriate representation for the "genes" or random solutions. 2. Create an objective function to optimise (also referred to as a "fitness" function) 3. Formulate appropriate crossover and mutation operators.

The basic flow of a genetic algorithm is as follows: a. Generate random population of solutions (the initial/seed generation). In this case, a solution to the problem is a route across the area of interest. b. Score each solution using the fitness function. This consists of similarity between the predicted RSS at each c. If fitness threshold or max number of generations has been reached, end d. Take the top n solutions and apply crossover and mutation operators to them e. Pass the top n solutions, as well as the resulting additional solutions from crossover and mutation, into the next generation, with additional random solutions to maintain a fixed population size p f. Go to step b with the new generation.

Representation of the population members

Encoding the problem in a manner suitable for operation with a genetic algorithm is crucial to the successful application of the Genetic Algorithm (GA) technique. Fortunately, the RSS geolocation problem presented in this document lends itself well to being encoded as part of a genetic algorithm. We can represent a gene in the formulation as the following:

Table 1 - Encoding genes as part of the genetic algorithm

Generation of new genes

The input to the geolocation algorithm is a time-stamped set of observed signal values, as well as a propagation model that allows the RSS at specific points on the ground to be predicted, as shown in the previous section. The simplest approach to generating random solutions to the geolocation problem is: • For each new member of the population: • For each timestamp in the observed samples: • Generate random location L within the area of interest

• For each receiver R

o Generate predicted RSS for R at location L • Store predicted RSS values, location and timestamp into the new member of the population. • Store the new member of the population, iterate until the desired number has been reached.

Using this method, we can obtain many randomly generated paths across the area of interest, tagged with the predicted RSS value at each timestamp/location combination. Using this method will result in a large number of "impossible journeys"; that is to say, routes

across the map which are impossible according to the movement model used as part of the input to the geolocation system. For this reason, it may be advantageous to use the movement model to either generate a "guided" random path (if the movement model is capable of generating random "legal" moves), or to do generate-and-test, rejecting unlikely or impossible routes.

Combination of genes

Of critical importance is the method by which genes are combined.

In traditional genetic algorithms, two parent "genes" or solutions to the problem are combined by randomly assigning parts from each of the parent genes to a new child gene. This approach can be taken with the geolocation problem, but will probably result in a large number of impossible solutions. A more sensible approach is to iterate over each timestamp, examining the randomly generated location L in both solutions and determining if it is possible to navigate between them according to the movement model. A graphical representation of this is shown in Figure 9. According to this figure 9, there is only a small area where Parent 1 (41) and Parent 2 (42) are close enough to be recombined (circle indicated by 43). If the genes were combined at this location, the solid and dashed paths would swap at the crossover point; the offspring would both have a very sharp turn introduced to their respective paths, which may or may not more closely match the path taken by the emitter.

Table 1 shows an example of how this might be implemented, assuming an emitter movement model that only allows movement of 1 unit within a given time window. Iterating over each time step, we identify that the emitter paths could be combined at location T2 (this location would correspond to the circled location 43 in Figure 9). Other locations are rejected, as the paths are not close enough to be accepted by the model. If parent genes are selected at random, using this method does mean that occasionally parents will not be suited to be combined (for example if they are on opposite sides of the area of interest, and never get close enough to intersect), resulting in no offspring. This case can be dealt with in a number of ways, but the simplest is to reject the pairing and try again with different parent genes, until the desired number of offspring has been generated.

Table 2 - Example Crossover Operator

Result Selection

After the genetic algorithm has been run for a number of generations (this number will be dependent on the propagation model, and propagation environment), the top scoring population should be chosen and returned as the result.

Implementing an enhanced movement model

Earlier in this document, an enhanced movement model is described. The following section contains some notes on how this could be realised in a practical system. In this implementation, we assume that we know the emitter is being transported in a car, travelling around a sub urban area. In order to create our movement model, we require an example data set; this can be generated by driving a vehicle round a similar area, and recording a time-stamped GPS trace.

Generating Summary Statistics 1. Gather movement trace of example emitter, recording location and time every / seconds 2. Divide trace into segments of duration n seconds 3. Calculate summary statistics for each segment 4. Summary statistics are used as input to movement model.

The optimal segment length n is a parameter that should be adjusted according to the environment, and the predicted behaviour of the emitter; a highly dynamic emitter may be tracked better using a shorter segment window. The best value for this should be determined through experimentation or modelling. In an ideal situation, / should be set to the same interval as the RSS readings are taken by the receivers in the system; however it would be possible to adapt the system if this were not possible, interpolating readings to deal with any discontinuities in the data.

The summary statistics on the data can be calculated using a software package (for example those included in Matlab®, R, Python®). The technique often calls for the magnitude of speed or direction changes (or deltas); these can be calculated by the following: sse^s-Lfeitaj. =---- i j

Where: + *s a function that returns the distance between two locations * is the predicted location of the emitter at timestep £ * ί is the sampling interval

Bearing deltas can be calculated in an similar manner: . hwara^^i) freeri»^^) i δβΏπηβΡβΙία» = i---

j I

Where: * bear/ngfplzp2/ is a function that returns the bearing from location pi to location p2 » Is the predicted location of the emitter at Umestep, £ » i is the sampling interval

Noting that in this case we take the absolute value; unlike speed deltas, we assume that we do not care which direction an emitter is turning, only the magnitude of the change in direction.

Applying summary statistics to score model

Once the summary statistics described earlier have been computed for the example emitter, they should be stored in a table or database. The goal of this is to provide a dataset to allow comparison between the predicted emitter paths generated by the geolocation process, and the paths produced by the example emitter. The basic principle behind this is to search the database for the most similar set of summary statistics, and use the difference between that and the statistics of the predicted emitter path as a metric for scoring the solution. An emitter that behaves in a similar manner to the one used to generate the database will produce similar statistics. 1. Divide the predicted movement path into segments of duration n seconds 2. For each segment statistics P in the predicted path: a. For each segment statistics E in the example database: i. Calculate difference between P and E statistics as described above ii. If difference is less than best difference so far, record the magnitude of the difference 3. Output of the process is the list of minimum differences for each segment in the predicted path

Calculating the difference between statistics will depend on which of the statistics outlined earlier was implemented; it will be necessary to apply a weighting to each statistic, this should be determined through experimentation or simulation as it will vary depending on the environment that the system is deployed in.

There are a number of methods that could be used to attach an overall score to a path; the simplest is to simply take an average of all the segment differences produced in the algorithm described

OPTIONAL EXTENSIONS TO THE SYSTEM

Extension 1 - Sanity Check

The performance of the system may be enhanced further by including a "sanity check", in the form of a basic check using a propagation model.

It is possible to formulate the geolocation problem in a geometric fashion, using the principle of intersecting circles. In this scenario, we assume that we have three or more receiving nodes, observing a signal transmitted by an emitter that we wish to geolocate. We begin by creating a table of observed received power levels, for example:

Table 3 - Observed received power levels

As we do not know the transmit power of the emitter, it is not possible to convert these observed power levels into absolute distances. However, we can make the assumption that the ratio of

the received power at two nodes is proportional to the distance between them. This can be expressed by the following equation:

Equation 11 - Geolocation using power ratio

Where the following definitions apply: • + is the x coordinate of the first receiver (e.g. units of metres) • Sis the y coordinate of the first receiver (e.g. units of metres) • % and S are the x and y coordinates of the emitter . ancj are the received power levels observed at receivers 1 and 2 respectively.

It is possible to generate these equations for each pair of receivers; the geolocation problem then becomes one of finding the values of % and % that satisfy the system of equations. Unfortunately, due to the characteristics of propagation modelling it is very unlikely that a perfect intersection will be found between three or more sets of results. The problem is therefore more likely to be solved by the application of a least-squares-error type optimiser, searching for the values of and that yield the smallest error.

Extension 2 - Tracing multiple emitters

The system can easily be extended to trace multiple emitters, provided the emitters are distinguishable from each other in some

manner. In the GSM® protocol for example, mobile phones send a unique signature which would distinguish them from other mobile phones in the area of interest which are also transmitting at the same time. Alternatively if the multiple emitters are transmitting simultaneously, each at a different frequency, the signal receivers would be able to differentiate these signals and attribute them to specific emitters. In this manner, multiple emitters can be traced simultaneously. Each emitter being geolocated represents a different optimisation problem. There is no limit on the number of emitters that can be geolocated except for computing power. Processing power required to locate multiple emitters would scale proportionally to the number of emitters being geolocated.

Different EMMs could be applied to different emitters being simultaneously geolocated, e.g. geolocating someone on foot and simultaneously geolocating someone in a car.

Extension 3 - selecting the best-fit EMM where nothing is known about the emitter.

Another optional enhancement would be selecting the optimal (best-fit) EMM to apply if information is not known about the emitter.

If the mode of transport of the emitter is unknown, it may still be desirable to operate the geolocation system. For example, an emitter may be carried on foot, or in a vehicle, and this information may not be available to the operator of the system, even though movement models for both pedestrian and vehicle born modes of transport exist. In this situation it will be necessary to select the optimal movement model from multiple alternatives that are available. The technique outlined below is a solution to this problem. The scenario assumes that an RSS similarity metric and optimisation technique has been selected. The algorithm below is applicable to the movement model described in the preceding section, although could be adapted to suit other models. 1. For each candidate movement model profile a. Generate a geolocation result using the current candidate movement model profile b. Store the geolocation result, similarity metric output, and movement model output as described previously 2. Select the geolocation result with the highest combination of RSS similarity metric score and movement model score as the result.

It will be necessary to apply a weighting/scaling factor to the RSS similarity and movement model similarity described in Step 2 of the algorithm above. This should be determined through experimentation or simulation, as it is likely to vary depending on the deployment scenario (in some situations matching movement model and predicted movement profile may be more important than matching predicted and observed signal strengths, and vice versa).

Example Results

This section provides a typical result from the system; it is based on a dataset gathered as part of the development of a trial. Five receivers were distributed around the test area, approximately a 2.4km by 1.5km area. These receivers consisted of an Ettus Research Universal Software Radio Peripheral (USRP) and a low power PC running Linux® and the GNU® Radio software.

The signal source used for the trial was a Rohde & Schwarz® SM300 signal generator configured to output a carrier wave at 453MHz and 872MHz. RF modelling was done using the Prophecy RF modelling tool after importing off the shelf LIDAR data purchased for the purposes of the trial.

The signal source was installed into a vehicle, and driven around the test area. The test area contains an extensive network of roads and buildings, representative of a suburban setting.

The data captured was time-stamped complex I/Q data at each receiver, which was post-processed using further GNU Radio scripts to extract received signal strength readings. Time-stamped GPS traces were recorded from the test vehicle.

The GPS traces were combined with the time-stamped RSS values from each receiver to produce the dataset. The geolocation technique was implemented as a Java® program and run on a number of journey segments extracted from the dataset.

Figures 10 to 12 provide some example results to give an idea of the output from the algorithm.

By contrast Figure 13 demonstrates the output from a basic implementation of an typical RSS geolocation algorithm that simply finds the closest match between predicted and observed RSSI vectors; as this shows, there is a lot of ambiguity in the results using traditional RSS geolocation methods whereas the improved method provides a significant step forward in accuracy (the improved geolocation results arising from the invention are confined to the box shown in Figure 13 and can be seen expanded in Figures 11 and 12).

Glossary

Area of Interest - Refers to the area over which the geolocation system is designed to operate. Emitters are assumed to be inside the area of interest. Receivers do not have to be placed inside the area of interest, but typically are. COST Model - Propagation model that extends the Hata model, developed by the "Cooperation europeenne dans Ie domaine de la recherche Scientifique et Technique" DF - Direction Finding. A geolocation technique based around determining the line of bearing from which a radio signal is received at a receiver. Examples include Angle of Arrival (AoA) and Pseudo-Doppler techniques.

Emitter Movement Model - A model of motion for a particular mode of transport that takes a time-stamped set of geographic points and outputs a numerical estimate proportional to the likelihood of that sequence of actions having been taken by the mode of transport.

Evolutionary Optimisation Technique - A class of biologically inspired optimisation algorithms. Such algorithms are frequently used to find near-optimal solutions in cases where finding an optimal solution is impossible or impractical.

Freespace Model - A propagation model based on the principles of the propagation of radio waves in open space.

Genetic Algorithm - An example of an Evolutionary Optimisation Algorithm. GNU Radio - An open source set of digital signal processing (DSP) libraries. GPS - Global Positioning System. Geolocation system based around on-device Time Of Arrival observations of a network of satellites. GSM® - Global System for Mobile Communications. The most common of the second-generation (2G) mobile communication standards.

Hata Model - A commonly used propagation model based on a set of equations formulated by Hata fitted to curves obtained from measurements taken by Okumura I/Q Data - Sampled data containing both in-phase and quadrature components of a signal.

Java® - A widely used general purpose programming language.

Kurtosis - A measure of the "peakedness" of a probability distribution; i.e. if its standard deviation is due to a few large deviations from the mean, or many small ones. LIDAR - "Light Radar". A laser ranging technique commonly used to create high-resolution terrain maps. MATLAB® - A widely used high level software language designed for development of engineering applications

Okumura Model - A model for land mobile radio systems based on measurements taken by Okumura in Japan.

Propagation Model - A model that predicts how radio waves are attenuated or otherwise influenced by environmental factors.

Python® - A widely used general purpose scripting language. R - A programming language and environment focused on performing statistical analyses

Ray Tracer - A propagation model that uses knowledge of structures in the environment to accurately predict propagation effects, for example through calculation of reflection, diffraction and penetration losses. RSS - Received Signal Strength. A measurement of the power present in a received radio signal.

Signal Data - Data pertaining to a signal of interest; examples could include received signal strength, frequency, duration, modulation scheme or technology.

Software Defined Radio - A radio system where the majority of signal processing (e.g. mixing, filtering, modulation) is undertaken in software rather than in specialised hardware. TDOA - Time Difference of Arrival. A geolocation system based around a network of synchronized receivers. The arrival time of a detected signal is recorded at multiple receivers and sent to a central point for processing. TOA - Time of Arrival. A geolocation system based on observation of the time of arrival of multiple transmitted reference signals by a device. USRP - Universal Software Radio Peripheral. A software defined radio developed by Ettus Research

References

Genetic Algorithms - Machine Learning, Tom Mitchell, McGraw Hill, 1997, ISBN 0070428077

Depth/Breadth first search - The Algorithm Design Manual , Steven S. Skiena Springer Publishing Company, 2008, ISBN 1848000693

Claims (9)

1. Claims 1) A means to estimate the geolocation of a moving signal emitter/receiver comprising:- a) A plurality of receivers or transmitters for receiving or transmitting signal data to or from the moving signal emitter/receiver; b) Collecting the signal data at the signal receivers or at the moving signal emitter/receiver at multiple points in time and generating a timestamped Received Signal Strength (RSS) vector for each of these points in time; c) means to determine the impact of features in an area upon the RSS, generating predicted RSS over said area, and using this to generate predicted Received Signal Strength (RSS) vectors at more than one location within said area; d) Estimating probable positions over time for the moving signal emitter / receiver based upon steps b) and c); Characterised in that e) there is provided means to generate a series of candidate routes denoting the moving signal emitter I receiver movement; f) using an Emitter Movement Model (EMM) to estimate the probability of any of the series of candidate routes being taken; g) means to estimate the correct Emitter Movement Model to apply to the moving signal emitter/receiver; h) means to compare each timestamped Received Signal Strength (RSS) vector, each of which corresponds to a location within the area, with the predicted Received Signal Strength (RSS) vector corresponding to the same location within the area, to provide a similarity metric; and i) means to find the most probable of all of the candidate routes to determine the most likely route taken, by finding the product of the probability of the route being taken (in f) with the summation of the similarity metrics (in h).
2. 2) A means to estimate the geolocation of a moving signal emitter/receiver according to claim 1 whereby each timestamped Received Signal Strength (RSS) vector is collected in absolute form.
3. 3) A means to estimate the geolocation of a moving signal emitter/receiver according to claim 2 wherein the signal data is collected at the signal receivers and the plurality of transmitters or receivers comprise a first transmitter or receiver for receiving or transmitting signal data to or from the moving signal emitter/receiver and other transmitters or receivers for receiving or transmitting signal data to or from the moving signal emitter/receiver and each timestamped Received Signal Strength (RSS) vector from the first receiver or moving receiver/emitter is stored as a figure relative to the timestamped Received Signal Strength (RSS) vector received at the other receivers.
4. 4) A means to estimate the geolocation of a moving signal emitter/receiver according to any preceding claim wherein the signal data is collected at the signal receivers and the timestamped Received Signal Strength (RSS) vector(s) include data acquired by the plurality of receivers at a plurality of frequencies.
5. 5) A means to estimate the geolocation of a moving signal emitter/receiver according to any preceding claim wherein the signal data is collected at the signal receivers and the timestamped Received Signal Strength (RSS) vector(s) include data acquired by the plurality of receivers simultaneously from multiple unique moving signal emitter/receivers.
6. 6) A means to estimate the geolocation of a moving signal emitter/receiver according to any preceding claim whereby a plurality of EMMs are used.
7. 7) A means to estimate the geolocation of a moving signal emitter/receiver according to any preceding claim whereby the moving signal emitter/receiver being geolocated is a mobile phone.
8. 8) A means to estimate the geolocation of a moving signal emitter/receiver according to any preceding claim whereby the means to find the most probable of all the candidate routes to determine the most likely route taken comprises an evolutionary optimisation technique or a breadth I depth first search algorithm.
9. 9) A means to estimate the geolocation of a moving signal emitter/receiver according to any preceding claim comprising a geolocation system capable of geolocation of the moving signal emitter/receiver.