SE503515C2 - Detection and prediction of traffic disturbances - Google Patents

Abstract

The invention relates to a method for detection and prediction of incidents and traffic queues formed by overloading. This is done in real time with use of sensors in a road network. Predictions are used also to reach a faster and more reliable detection. Sensor measurements are also used in the process, where the comparison with expected values are used for successively updating stored parameter values for the involved algorithms. By this, the system can succeedingly adapt itself for changed situations. The strong traffic variations, that are naturally occurring at short time intervals are treated with the use of noise-based methods. By this, there are formed distribution related measures as e.g. the standard deviation, which can be estimated from measurements, and submit a base for estimating probabilities for deviations of a certain size, e.g. related to the standard deviation. Automatic incident detection (AID) is based on determination of the desired false-alarm rate, and the related threshold level. The method includes accumulated measurements. Faster and more reliable incident detections are received with the use of the invented prediction process method.

Description

503 515 The distance between sensors A and B can also be increased and the number of sensors reduced. The requirement here is instead that it is possible to make reasonable predictions. Roughly speaking, however, it can be said that even "weak" predictions should mean an improvement, for example, even if one could not predict exactly 0 cars in the example above, then in any case a prediction of a decrease in fate could mean a less measured deviation, than if you do not know anything, and thus less risk of false alarms. - 2A. Traditional methodology.

Another disadvantage of traditional methodology is that it is so much based on experimenting with doctrinal formulas with balanced parameters. It is often difficult to understand why one way would be better than another, and it is also difficult to verify. Nor is it known whether it is possible to improve the method, through more attempts and triumphs. - 2B. The excitement.

In the present invention, the understanding of the traction processes is utilized in a very direct way. An important part is the understanding of strong traversations. These can be considered as the result of stochastic processes, and if you measure, for example, fate in a sensor, you experience this as noise. With short measurement periods, you get relatively strong variations around a given average value. By utilizing knowledge about "noise", one can understand to use the information in these "noise variations", and not just see it as something that spoils the possibilities of making simple detections of incidents.

Through measurements you can obtain mean values and standard deviations and from theory and measurements you can create approximate distribution functions, ie you know statistically a lot about the track variations. Assuming as an example a normal distribution, a measured standard deviation can provide information about the probability that a variation will be greater than a given value. You also get an understanding of what it is that you do not know anything about. (And that it never pays off with never so long and parameter-filled formulas according to traditional methodology.) In the present invention, the knowledge of probability for deviations of a certain order of magnitude is used to set thresholds, which thus gives the desired false alarm risk. It may also be the case that a deviation due to an incident is not large enough to reach the threshold. Then you can wait until you get the next measured deviation and see if these two together are so large that the probability requirement is now met, ie that you now reach the corresponding threshold. This can be repeated successively, and if the natural variations are very large, the threshold will be high and incident your incident-caused deviations are required for them to reach above their threshold. However, time passes, the measurement periods are needed, and incident detection should be done very quickly to prevent serious secondary effects. In the recovery, for example, the threshold can be set automatically so that a minimum of extra measurement periods need to be used.

For this reason, it also becomes important to keep down the natural trunk variations, and this is done as stated above in the invention, in that it is not the variations from a mean or previous value that are taken into account, - but it is the smaller deviation between predicted value and measured , which determines the threshold level. Thus, the threshold level can be lowered significantly without the risk of falsification being increased and an incident-caused deviation then more easily reaches above the threshold and the incident is quickly detected. - 3A. Traditional methodology. 503 515 A third disadvantage of the traditional methodology is that it is difficult to transfer from one situation, where with “trial and error” it was finally made to function acceptably, to another situation. It may apply geographically, for example to another road section, where entrances, road crossings or the number of lanes give others the traffic situation. This may apply to changes in measurement times or other parameters. This can be a very time consuming and resource consuming effort. Namely, one needs to check the trajectory in parallel with other accurate measuring means in order to have a result to compare with, so that parameters in the formula can be changed towards a better correspondence with reality. Changes in the traffic situation may also mean that the process needs to be redone to get new, better parameters to fit into the formula. - 3B. The invention.

In the present invention, much of the adaptation can take place automatically. In addition, the start values can be selected well from the beginning. Regardless of the set-up, this current deviation is measured, and corresponding statistical measures are obtained, eg the standard deviation of the deviations. From these measurements, the respective threshold value can be set automatically, and the method starts to generate incident detection, which the operator can check the accuracy of. Because the method constantly continues to measure deviations, the statistical parameters can be gradually updated and adapted to changes in the traffic situation.

Traffic disruption and queues. Congestion of the road network, even if it only lasts for a short period of traffic, is enough to cause the traffic to break down and queues to grow. These queues can then be maintained by a lower traffic fate, as the capacity of the road network usually decreases as a queue is formed. If, for example, at the same time as a traffic peak on the motorway, a traffic peak occurs on a driveway, so that the cars do not fit, then the cars slow down during the attempt to weave the two fl together. The speed can then be reduced to very low values and the distances of the cars become short with the result that the fate is low. It can be significantly lower than the maximum fl fate, which can be kept below free fl destiny fi k, and which constitutes the capacity of the road link.

The result of congestion is thus that queues are formed, which reduces the capacity of the roads. In our large cities, this occurs daily during rush hour traffic in the mornings and afternoons, which means that the road network has its lowest capacity, when it needs to be at its largest.

It is thus very valuable to be able to lead traffic so that traffic breakthroughs are avoided.

It is especially important to prevent traffic breakthroughs in critical places, where queues lead to serious disturbances on one or your major traffic routes. This can be prevented by means of the present invention.

A key function is predicting traffic breakthroughs and queuing. Through prediction, a time margin is obtained before the predicted problem actually occurs. This time can be used to introduce measures, which prevent the problem from occurring in reality. Even when detecting that a queue has arisen, it is interesting to use prediction. If, for example, free fate was predicted and a queue still occurred, the sensors give values that show the actual traffic situation (queue). The discrepancies between the predicted free des values and the measured values can thus be used as an indication that a queue has arisen.

In this description of the claimed invention, terms other than "predicted" are sometimes used, such as the word "expected". In general it can be said that if a method contains 503 515 that a measured value is to be compared with another, previously known, "corresponding value", then the assertion of "corresponding value" often means a future reference, even if the value was obtained only from historical values Therefore, the term predict will also be used, including estimates, which are not direct predictions, but serve a similar purpose, for example, the comparative value may consist of an average value or a mean value plus a value based on standard deviation, historical calculated value, etc. produced, the purpose is nevertheless that this constitutes a form of expected iron value, with which the measured value can achieve criteria for detecting a queue.

Thus, the expected value has a forward-looking function against the measured value, and is estimated in an equivalent process as a prediction, even when the expected value is calculated afterwards, ie after the measured value has been obtained.

According to the invention, queue detection can also be done when queues are formed on road sections between sensors. This also applies to the use of video sensors, IR sensors, radar and similar sensors, which, for example, with their image can cover a longer distance than the meters covered by traditional loop sensors. However, the "range" with video sensors is in practice much shorter than the distance you "can see". The camera's limited position in height means that, for example, a bus can obscure a long distance with cars. Video sensors with a distance of 0.5 - 1 km can therefore only guarantee coverage in their immediate area, and most of the intermediate distance may be handled in the same way as with loop sensors.

Detection can be performed at both downstream and upstream sensor. At the upstream sensor, the gender is detected by the gender being within the sensor's direct measuring range. The characteristic of the queue is that the track is dense and moves at a lower speed than with free flow. It is known that when the när fate approaches the road's capacity ceiling, the speed drops, for example at a motorway entrance with a speed limit to SEK 70 / h, the speed can go down to 55 km / h due to traffic congestion. With even denser traffic, traffic collapses into a queue, which can have even lower speeds. According to the invention, the latter traction condition can be checked by measuring for at least two measuring periods. It turns out that when there is a queue due to a traffic collapse, speed and flow are phased, so that both fate and speed increase and decrease together. On the other hand, when there is heavy traffic, including moving queues at higher speeds, speed and fate change in opposite phase. Using this method, it is also possible to define at which breaking speed dense trajectory typically turns into a queue collapse.

Measurements in Gothenburg by the inventor, showed that the breaking speed was at 55 kni / h. This was well reproducible at the typical speed limit of the road network, 70 km / h.

There are other definitions of queue, for example one where already two cars can be considered to form a queue if the time gap between them is less than 2 seconds and the car behind has a higher demand speed.

Queues and queues also have different processes on country roads with one lane versus two lanes and compared with motorways. The queues that are most interesting for the current patent are those that arise on motorway-like entrances and exits to larger cities.

From a tracing point of view, the essential queues are those that create major problems. Therefore, small groups of cars driving close to each other are considered dense traffic. Even longer packages with cars traveling at slightly reduced speeds compared to the free-speed speed (usually the signposted speed) are considered here as dense traffic. Usually such car packages are characterized by the front of the package moving forward along the road (so-called moving queue). At speeds above the breaking speed, the traffic in such a package is characterized by high fate and reasonably high speed, which is why homogeneous (calm) driving in such a package does not constitute a direct traffic problem.

In the vicinity of cities, however, it is crowded with av- and pafaiter, so calm dense traffic is rare. Instead, the traffic is characterized by lane changes, “weaves”, which rather cause the dense traffic to collapse, resulting in low-speed queues in the unstable queue council.

For motorway traffic, the so-called stretching system used. These systems are traditionally based on setting up variable speed signs at intervals of, for example, 500m, along a motorway section. Sensors measure the traffic situation, and when the traffic becomes noisy, varying speed, etc., the system lowers the speed limit over a number of intervals.

The intention is that the traffic will be calmed and the speed homogenized to the now signposted lower speed. There are reports that claim that significant improvements have been made regarding both accident reduction and increased passability. A problem with traditional methods is that when the chaos occurs, the traffic collapse is already on its way, and when measured values and detection have been obtained, the collapse is often a fact. Various attempts are made to improve the situation, such as measuring in very short periods to get information quickly.

In the current invention, traffic is instead predicted gradually, and when the probability of collapse exceeds a specified value, the corresponding speeds on the signs are reduced.

This creates time margins to avoid traffic collapse, and interventions in traffic can also be reduced. The method is the same as that used for queue and incident detection.

The current invention can also be used in the control of oncoming traffic, eg control of “ramp metering”. In the vicinity of cities, oncoming traffic and exit traffic is a serious source of disturbance as described above, where congestion and traffic collapse often occur to difficult queues. be used for reducing the oncoming traffic on the ramp in question, and for reducing the traffic on the motorway, by, for example, reducing oncoming traffic on ramps upstream of the current approach.

Prediction of traffic breakdown at an access road can be based on measurements in upstream sensors, eg a sensor on the main road and a sensor on the access road. Measurement of traffic in each sensor can be used to predict the traffic some time later, equal to the travel time, at the joint weaving point. By matching or synchronizing the measurements, for example, occasions when simultaneous traffic peaks reach the driveway can be predicted. The predicted fl fates are compared with threshold values for prediction of overload. A way to set the threshold value for the main route is illustrated as follows. The weaving capacity Cv = Co - a * Ie, where Co is a constant and Ie is fl the fate of the driveway. Factor a shows that the capacity of the main path is not determined by a simple sum of the two fl fates. Both Co and a should be calibrated for the current approach. Examples of values for a two-lane motorway are Co = 4200 / h and a = 2.5, ie a differs considerably from the value 1. This is also evident from the value of a at an exit, where for example a = 1.5. In both cases it is the weaving that gives the smaller total capacity at resp. driveway. This formula has been shown to fit well down to small drive-in fates. When the traffic has broken down, some other conditions apply. In the case of an established queue, for example, every other weaving often takes place, which gives a corresponding relationship between the traffic on the main road and on the driveway.

As queues grow so that they reach the upstream sensor, algorithms, such as those mentioned above, can be calibrated, and Co and a, respectively, can be successively updated. Algorithms that describe queue growth and queue settlement can also be updated.

Queue growth is determined by the difference between the fate behind and in front of the sex. The flow in front of the queue can, if necessary, be estimated from a model for queue settlement in the front of the queue. An example of a simple formula is Iaq = b * (g) '°' 5, where g is the gap between the cars in the queue front and b is a 503 515 constant, the value of which can be stated approximately and updated by measurements. Iaq can also be determined from Iaq = (Va - Vq) / (Da - Dq), where V is speed and D is the period distance of the cars, and a indicates parameters in front of the queue and q in the queue.

Together with the relationship I = V / D, the fl fate in front of the sex and fl the fate in front of the sex can be determined, and with information about fl the fate and the current speed behind the sex, the growth and settlement of the sex can also be determined. The queue settlement formula applies to many common situations, and the gap g can be obtained from typical relationships between gap, flow and velocity at queue conditions.

Determination of probability for queues and incidents.

When predicting an event, it is not always most interesting to assess whether it is most likely that the event will occur, ie if the probability is over 50%. If the risk of queuing is 30% or the risk of an accident is 10%, this may be sufficient for measures to be taken to prevent the events from occurring, ie even though it is most likely that there will be neither a queue nor an accident. Below are examples of how according to the invention one can work with probability determination.

A very common distribution function in statistics is the normal or Gaussian distribution. If this is assumed to apply approximately to the car traffic on the relevant part of the road network, the function can be calibrated from measurements of the variance in the traffic around the mean value. The probability of obtaining a certain value can be calculated or usually taken from a table. Depending on how the detection is performed, modifications to the distribution may be needed, or adaptation using other distribution functions.

The Rayleigh distribution is, for example, interesting for envelope detection and filtered noise deviations. As an illustration of the invention, we use an approximate distribution of "noise" deviations according to P [y (t) 2x] = exp (-x2 / o2), ie that the probability that a deviation y (t) is greater than a given value x is exppxz / øz).

We then get that the probability that during a measurement period there is a deviation greater than the standard deviation is exp (-1) = 37%. We can also ask how large the deviation needs to be for the probability to be as small as, for example, 104. From 104 = exp (-9.2) it is obtained that x = 30. This also means that during 104 measurement periods there is a 37% probability of finding a deviation greater than this value. If a large deviation indicates that it may be an incident, then a threshold setting of x = 36 means that the risk of falsification is 104.

In the example above, x and y can be fl fates. In that case, incidents are expected to give rise to reduced flows in front of the sex, and thus only one-sided deviations need to be taken into account, whereby the probability values are half of the above.

When accumulating fl your measurement periods, the noise deviations are added quadratically, ie after n periods, o (n) = (n) ° '5 * o' (1). The signal is added linearly, ie as a comparison s (n) = n * s (1).

This gives an integration effect s (n) / o (n) = n ° '5 * s (l) / o (l).

At the false alarm risk 104 it is now obtained that the threshold related to a measurement period can be lowered at accumulated measurement periods according to (9.2 / n) ° '5, ie x = 3 * n' ° '5 6.

In a practical example, for example, the fate of cars during a measurement period of 30 sec. on average 16 cars and the associated G-value is 4 pcs. If an incident blocks a lane and the queue formation reduces fl the fate to 8 cars, then the deviation becomes 8 cars = 20, and the condition for obtaining 503,515 Irish excess in faiskiarmsfisk ma io * is then an 20 is Siam than (9.2 / n) ° ~ 5 o.

The number of measurements then needs to be greater than 9.2 / 4, ie 3 pcs. Had the distribution instead been linear in the exponent, ie exp (-x / o), more measurement periods would have been needed.

According to the invention, one can instead use the deviation between predicted and measured value for the distribution function. In this case, the deviations can be reduced considerably.

Assume, for example, that 70% of the deviations can be predicted so that o '= 30% of 4 = 1.2 pcs. Then it is required that the signal is greater than 9.205 * 1.2 = 3.6 cars. With an incident-caused deviation of 8 cars, a reliable incident detection is obtained already in the first measurement.

As there are many sensors in a network, it is even more important to keep down the risk of false alarms. If you want to limit the false alarm to one per day with 100 sensors, l0O * 24 * 60 * 2 periods are available, and the false alarm requirement is 3.5 * 106.

According to reports, traditional methods have problems with long detection times, which are also illustrated above. As shown, the invention provides very large time savings.

Below are some examples of the use of the methods of the invention in the traffic control process.

Traditional methods for controlling traffic at, for example, a driveway, involve an isolated control of the traffic point by point on the given ramp.

With the invention and its use of prediction, time margins are created for controlling the tracks coherently in a network. This is a big step, as traffic is very much a network function. Problems arise in cramped sectors, - but problem solving should be network-based, otherwise the problem is often just moved to another point.

Closely associated with network control is the control of traffic between different routes. If when predicting traffic on a road network, traffic problems are predicted, for example congestion, this problem can often be avoided by taking measures in time, eg traffic can be diverted to another route, whereby the traffic volume on the original route decreases to a level below overload.

Route control can, for example, take place with the help of "VMS", adjustable signs. The message can, for example, contain information about different degrees of problems on the given route. The bigger the problem, the more likely you are to choose an alternative path. Through sensor measurements in direct connection with motorists' route selection, rapid feedback is obtained on the proportion of road users who chose the alternative route. This measure is also used to update the strength value of the current displayed message, so that the system gradually stores a current measure of the strength of each message. Thus, the system can pre-select messages that suit the proportion of motorists who wish to choose a new route.

The invention includes that one can predict the outcome of measures. This is important because one should not choose a measure that leads to new problems. Calibration and updating are obtained by successively measuring the consequences of the measures and adapting the stored strength values for the measures. In this case, a slower update is preferably chosen, so that only part of the change has a direct impact.

The invention is also suitable for controlling so-called "park and ride", for example parking the car and taking the train or bus, - where the control information is partly based on predicted problems in the 503 515 road network. Another area is the control of departure time, for example, information about traffic problems can persuade road users to choose another means of communication or to postpone their journey.

Claims (15)

503,515 Patent claims. 1. l. Method of determining traffic jams in a network by means of sensor information from several sensors, where queue formation on a route between two sensors can also be detected by means of at least one of the upstream and downstream sensors, characterized by; that with the downstream sensor the resulting difference in fl fate at the queue is detected compared with non-queue, and where the difference in fl fate in the queue case is determined by the fl fate in the front of the queue; and that with the upstream sensor a queue is detected, when the queue has grown to the position of the sensor, and that the queue situation is determined by the speed reduction from the free fall, preferably by measuring at least one of the parameters fl fate and traction density in addition to speed; and during prediction, queue formation is predicted with the help of upstream and possibly downstream sensors, in that fl fate relationships are determined and that these relationships are used together with current fl fates to predict queue formation in at least one narrow sector along the given route; and that the queuing process consists of two steps, where the first step consists of traceability breaking that starts the queuing, and step two consists of the growth and settlement of the gender; and that prediction of the first step is performed by means of prediction of fl fate on the route and where degradation is determined by fl fate including ev. exits and entrances exceed the corresponding threshold values; and that successive calibration and updating of input parameters in the queuing process is performed by comparing predicted values with corresponding values measured in the sensors and thereby the parameters are corrected successively by preferably slow, filtered or scaled parameter change. 2. A method according to claim 1, characterized by; that when the flow in the downstream sensor is predicted, this fl fate is compared with the corresponding later measured fl fate and that in this case a meat detection process uses the difference between predicted and measured values, possibly accumulated over fl your measurement periods, to determine if a queue has formed; and that in one or both cases, queue or non-queue is predicted and compared with measured values. Method according to patent claim 1 or 2, where the deviations between expected and measured values form stochastic distributions depending on the current traffic situation, eg free d desert or traffic containing queues, and where deviations greater than the standard deviation o 'become increasingly unlikely, if they are not expressions of a corresponding change in traffic, characterized in that individual deviations or accumulated deviations are related to the said distributions and when the deviations are greater than the chosen measure, this is assessed as a criterion for a corresponding change in traffic to occur. 4. A method according to claim 1, 2 or 3, characterized by; that in the meat detection process the probability of queuing is estimated, and that this is done by calculating measurements of the deviations, eg standard deviation, and these measurements can be related to an assumed or by means of measurements determined deviation distribution, and can at measured deviation or accumulated measured deviation, on the probability of queuing is obtained. 5. A method according to claim 1, 2, 3 or 4, characterized by; that incidents are detected using corresponding principles, which apply to meat detection with a main focus, that when a queue has been detected and the probability is small that it has been generated by other than incidents, this is a reason for assessing that an incident may have occurred; and applies when traffic flows are low that an unexpectedly stationary queue front was probably caused by an incident, and that when the traffic conditions are such that queues probably arose from congestion, additional information is required to determine that the cause was an incident, such as an almost completely blocked traffic during certain time, where congestion results in a queue with limited, but still significantly greater passability; Method according to patent claim 5, characterized in that the standard deviation or equivalent measure of deviations between predicted or otherwise expected values and corresponding measured values is used multiplied by a factor q to give a measure of the probability of single or accumulated measured deviations with the value (q * the deviation measure) represents the assumed traction situation; and q is determined by the employed distribution; and q can be defined as a function of the probability or stored in tabular form related to the corresponding probability value. Method according to claim 6, wherein there is a desire to keep the number of false alarms for incidents below a given false alarm level, characterized in that the corresponding probability of false alarms is determined and associated values of q are used to determine the threshold level, eg q * o; and comparing the deviations with the threshold value to accept a possible incident detection; and when accumulated deviations are used, these deviations are compared using a q-value which depends on the number of accumulated values; and this can be done, for example, by an approximate method, where the accumulated value is divided by the root from the number when compared with the use of such a q-value, which can then be considered substantially independent of variations in the number of accumulated values. Method according to patent claim 5, 6 or 7, characterized in that the traffic is also predicted as if an incident has occurred and that even measured deviation from this condition is used in assessing whether an incident has occurred. Method according to patent claim 7 or 8, characterized in that meat detection can be treated in a manner corresponding to that described for incidents. Method according to claim 1, characterized in that the determination or prediction of av fate on a given road section utilizes fl fate relations, which are connected to upstream sensors, eg a sensor on the main road and a sensor on an access road; and that prediction of the flow in the interconnection takes place essentially in the same time phase, synchronized, so that for example the prediction matches the interaction of two traffic peaks in the interconnection, and thus the higher risk of traffic disruption and queue formation, and, for example, the threshold value for the main road can be expressed approximately as Cv = Co - a * Ie, where Ie is fl the fate of the approach or exit and a is the associated factor and Co is a constant; and Co and a can be employed from experience values and possibly updated for each road link by comparing predicted values with the actual measured sensors. 11. ll. Method according to patent claim or 10, characterized in that the av fate in the front of a queue is approximated by b * (g) '°' 5 where g is the gap between the cars in the queue front and b is a constant which can be updated by comparison with measured values; and that by fl fate can also be so: 515 ll is determined from (Va-Vq) / (Da-Dq), where V is speed and D is the period distance of the cars and a indicates the situation in front of the queue and q in the queue; and that these expressions together with tra fi k fl fate, I = V / D, are used to determine the fl fate in the front of the gender and with information about the bakom fate behind the gender also the growth and settlement of the gender. Method according to one of the patent claims 1 to 11, characterized in that they are used in so-called motorway systems for controlling variable speed control. Method according to one of the patent claims 1 to 11, characterized in that they are used for controlling driveways, for example by means of "ramp metering". Method according to one of the patent claims 1 to 11, characterized in that they are used for route routing. Method according to one of the patent claims 1 to 13, characterized in that they are used for controlling traffic in a network.