ITUA20164219A1 - PROCEDURE AND SYSTEM FOR DETECTION OF VEHICLE TRAFFIC, CORRESPONDENT IT PRODUCT - Google Patents

Description

"Procedure and system for the detection of vehicular traffic, corresponding IT product"

TEXT OF THE DESCRIPTION

Technical field

The description refers to the detection of vehicular traffic.

One or more embodiments can find application in the detection of the average travel speed, for example of road sections.

Technological background

The calculation of the average speed with which a vehicle travels along a road section is a conceptually simple problem if one considers only the logical-procedural aspect, which concerns the association of the passages of the same vehicle entering / exiting the section and the numerical calculation of the average speed as the ratio between distance traveled and transit time.

However, when we consider "forensic" implications (for example of a sanctioning nature) and the requirements of accuracy and certainty of the measure that they impose, it is observed how the measurement of space and time become critical and such as to make the use of technologically simple solutions is problematic, especially on short distances traveled at high speed.

Automatically measuring the average speed of vehicles traveling along a stretch of road involves observing the same vehicle in points A and B (at a known distance) at different times, using synchronized devices.

Given to solve the systemic aspects relating to the estimation of the length of the stretch, the main technical difficulties are attributable to:

- determine with known accuracy both the position and the two instants of passage of the vehicles with respect to A and B;

- manage the inevitable competition of several vehicles, with the ability to distinguish them in order to calculate the average speed for each of them; this requires that the entire observation process and subsequent logical association be reliable.

Various automatic systems provide for distinguishing vehicles by means of sensors capable of providing an accurate representation of one or more discriminating parts of the objects themselves.

The most complete and immediate representation to be acquired, in particular in the external environment and with the vehicle in motion, is of the visual type, ie detected by optical sensors or (tele) cameras with acquisition of the relative images.

The 2D signal acquired by a camera can be processed by an automatic recognition system capable of providing a synthetic identification description of the object at a higher level of abstraction than its physical reality.

The plate is an example of a synthetic, distinctive representation of the vehicle, which contains much more detailed information in a simple and compact data (a string of text).

Documents such as EP 1 276 086 B1, EP 1 870 868 B1 and again EP 1513 125 A2 (substantially similar to EP 1276 086 B1 in this regard) use the conceptual approach described above.

For example, EP 1 276 086 B1 uses multiple sensors (cameras and microphones) and multiple automatic recognition systems applied to the related signals, described in a generic way, e.g. without addressing a topic such as calculating the size of a vehicle from images in which the vehicle appears in prospectively different positions A and B. Similarly, in relation to the microphone, the variation of the sound as a function of the punctual speed in the two detection points A and B does not appear to be taken into consideration, nor does it consider the existence of motive forces other than the internal combustion engine. The layout of the marks on the vehicle is also mentioned, information of great interest to establish the identity of the vehicle, but far from trivial to detect and match in the presence of inevitable variations in perspective and lighting.

EP 1 870 868 B1 foresees the use of only one sensor (camera) and one automatic system (automatic character reader, OCR) in A and B. In both cases, the "vehicle" object is described through one or more signals and related character recognition processes, with storage in symbolic form for a time necessary for their comparison at different instants. The automatic association of a vehicle in two different points and instants is therefore obtained using the textual syntheses provided at the output of recognition processes.

Such an approach can suffer from the fact that the automatic systems used are affected by probability of detection and probability of correct recognition, both of which are lower than unity. In practice, cases of non-reading or reading with one or more incorrect characters are certainly possible due to, for example, physical degradation or the presence of dirt that reduces its visibility. Although it can be certified by laboratory tests under ideal shooting conditions and license plate quality, the reliability of the license plate reading system drops in less than optimal conditions, for example in the event of snow, rain, fog. The small size of the plate and its typical low-altitude position with respect to the vehicle also reduces its visibility and exposes it to degradation of legibility due to additional interfering effects such as occlusions caused by other vehicles in the scene shot by the camera. Furthermore, the license plate recognition process consists of a sequence of concatenated sub-processes, such as: license plate locator, segmentation of individual characters, OCR (Optical Character Recognition), syntactic verification of the recognized sequence of characters, each of which is characterized by a probability of error such as to make the probability of recognition of the entire process less than unity; and the combination of independent processes at detection points A and B further reduces the likelihood of overall correct association. Basing the association only on the plate is therefore a definitely partial solution, prone to error. Finally, the use of an inherently 2D signal reduces the accuracy of estimating the vehicle's real-world position relative to the road as it passes through detection points A and B. In addition to the inherently reduced accuracy of a 2D signal in estimating the vehicle's position in the real world the absence of a clear metrological traceability of the two devices and of evidence of the maintenance of their set-up is important. In fact, a camera can trivially change its orientation angle and the focal length of the zoom lens, for example due to vibrations or mechanical drifts, thus introducing a localization error in the detection point, without however being able to give evidence of this variation on the 2D image of the vehicle in transit. Moreover, reading the license plate alone may not allow recognizing the vehicle class, which may be a factor that contributes to the determination of the relative and specific speed limit.

Document EP 1 997 090 A1 describes a system capable of providing both the 3D kinematic characterization of its motion (position, time and speed of passage) and the 3D representation of its shape. These data are obtained using in addition to the stereoscopic images also the different perspective positions that the vehicle itself assumes during its transit with respect to the stereo-cameras. The document EP 1997 090 A1 therefore uses the 3D and symbolic representation of each vehicle to describe it in its instantaneous passage, without however addressing the issue of obtaining combined representations of the same vehicle provided by devices installed in different places and collected at deferred times.

Purpose and summary

One or more embodiments have the purpose of overcoming the drawbacks outlined above.

According to one or more embodiments, this object can be achieved thanks to a method having the characteristics referred to specifically in the following claims.

One or more embodiments can be related to a corresponding detection system, as well as to a computer product that can be loaded into the memory of at least one processing device and comprising portions of software code for carrying out the steps of the process when the product is performed on at least a computer. As used herein, a reference to such a computer product is intended to be equivalent to a reference to a computer readable medium containing instructions for controlling the processing system in order to coordinate the implementation of the procedure according to one or more embodiments. The reference to “at least one processing device” intends to highlight the possibility that one or more embodiments are implemented in a modular and / or distributed form.

The claims form an integral part of the technical teachings administered herein in relation to the embodiments.

According to one or more embodiments (capable of satisfying the requirements of sanctioning applications), a vehicle traffic detection system can include (at least) two transit detection devices based on the use of stereoscopic cameras (hereinafter stereo - chambers) metrologically certified for space and time measurements, endowed with a state (intrinsic and extrinsic calibration) with the property of being constantly observable, placed at the entrance and exit of the road section to be controlled (of known length).

In one or more embodiments, these devices can be separately capable of providing, for each vehicle in transit, the 3D kinematic characterization of the motion (position with respect to a reference system on the road and time with respect to an absolute reference and speed of passage ) and the 3D representation of the shape, obtained by using in addition to the stereo images also the different perspective positions that the vehicle assumes during its transit with respect to the stereocameras. In one or more embodiments, such images can associate pictorial information (texture - Tx) with the 3D representation of the vehicle, thus giving rise to a representation that can be briefly defined as 3D + Tx.

In one or more embodiments, the (possible) use of several stereo-cameras in each control point, oriented in opposite direction and mutually calibrated, allows to cover the whole surface by adding the various 3D + type representations. Tx of the whole vehicle.

In one or more embodiments, the evaluation of the correspondence between incoming and outgoing vehicles for the calculation of the section speed can therefore exploit the wealth of three-dimensional pictorial and morphological information of the vehicle. In one or more embodiments, this information can encompass that undoubtedly distinctive part of the surface represented by the plate, taken up by both devices without however necessarily being automatically interpreted: in this way the overall solution is independent of any possible errors that may be committed. with synthetic data extraction processes.

In one or more embodiments, the accuracy of the estimation of the three-dimensional position of the vehicle with respect to the road section and of the time synchronization with respect to an absolute reference of the two devices facilitate the achievement of a high overall accuracy of the average speed measurement for each vehicle obtained through the relationship between the two primary quantities.

In one or more embodiments, the input-output association obtained through the possible correspondence checks, in "many-to-many" mode, that is, between the lists of vehicles transiting in and out in compatible times can prove to be more robust in the management of cases of non-detection (incoming or outgoing) compared to a more traditional "one-to-many" association technique, with the ability to provide a dynamic measure of the overall uncertainty of the process in order to minimize the associative error.

Brief description of the figures

One or more embodiments will now be described, purely by way of non-limiting example, with reference to the attached figures, in which:

- Figure 1 is a schematic representation of a possible context of use of one or more embodiments,

Figure 2 is a block diagram of a system according to one or more embodiments, e

Figure 3 is an exemplary diagram of the possible operation of one or more embodiments.

Detailed description

Various specific details are illustrated in the following description, in order to provide an in-depth understanding of various examples of embodiments according to the description. The embodiments can be obtained without one or more of the specific details, or with other processes, components, materials, etc. In other cases, known structures, materials or operations are not illustrated or described in detail so that the various aspects of the embodiments will not be made unclear.

A reference to "an embodiment" within the framework of the present disclosure is meant to indicate that a particular configuration, structure or feature described in relation to the embodiment is included in at least one embodiment. Therefore, phrases such as "in one embodiment" which may be present at various points of the present description do not necessarily refer to exactly the same embodiment. Furthermore, particular conformations, structures or features can be combined in any suitable way in one or more embodiments.

The references used here are provided simply for convenience and therefore do not define the scope of protection or the scope of the forms of implementation.

Figure 1 is a schematic representation of a possible context of use of one or more embodiments: the detection of the (average) speed V with which vehicles W (e.g. motor vehicles) travel a road section of length L between one (entry) point A and one (exit) point B.

For example, the speed V which is intended to be detected may be the average speed with which each vehicle travels along the stretch of length L, the detection being aimed at verifying e.g. that the speed V does not exceed a certain limit value, with the possible possibility of applying corresponding penalties to (drivers of) vehicles that exceed this limit.

In one or more embodiments, the detection can be delegated to a system comprising at least two (stereo) cameras SC1, SC2 capable of "framing" the vehicles W in transit at, respectively, an (entry) point A and of the (output) point B, supplying respective detection signals to two units U1 and U2, which can be configured to operate as exemplified in the block diagram of Figure 2.

As already mentioned, in one or more embodiments it is conceivable to use several stereo-cameras in each control point, oriented in the opposite direction (see for example the stereo-camera SC2 ', whose profile is represented with dashed line, associated with the stereo-camera SC2) with the aim of covering the entire surface by adding the various 3D + Tx representations of the entire vehicle W.

For simplicity and clarity of illustration, the present detailed description is made with reference to examples of embodiments which envisage the use of only two (stereo) cameras SC1 and SC2.

In one or more embodiments, the chambers SC1, SC2 can operate according to the stereometry techniques exemplified in EP 1 997 090 A1 and have the structural characteristics exemplified in the same document (see e.g. Figure 7), which makes it superfluous repeat here a corresponding detailed description.

In one or more embodiments (e.g. to allow the use for sanctioning purposes), it is desirable that the accuracy in the measurement of the average speed V can be contained within an overall relative error e.g. of 1%.

In one or more embodiments, this corresponds to providing that the maximum error of the two devices (SC1-U1 and SC2-U2), considered individually and independently, does not exceed a value equal to 0.5%.

In one or more embodiments, each device provides a measurement of position and crossing time for which 0.5% can be seen as the overall measurement error of the two quantities.

The absolute error of positioning estimation is linked to the length of the section and therefore the system is more sensitive to short sections.

For a section of even length L, eg. at 500 m, the above corresponds to containing the spatial error of each device must be contained within 2.5 m, a technically non-trivial theoretical result to be achieved and maintained, even assuming that the synchronization error is zero.

If we consider that the time synchronization error is also different from zero in reality (0.1 sec of error is an ambitious goal with the available technologies) and that even at only 100Km / h this time error is equivalent to one displacement of the vehicle of 2.7 m, the accuracy of the measurement of space and time represent a fairly stringent technological constraint.

In one or more embodiments, a representation of the single vehicle of the "3D + Tx" type (as defined above of the vehicle, that is, descriptive of its external three-dimensional surface, for example radiometric and chromatic), is able to provide a characterization of the vehicle such as to facilitate a safe and reliable association at different times and points of space.

It contains, with a high level of detail, the geometric and pictorial information identifying the vehicle itself, without mediations linked to the need to "condense" this information into a mere list of attributes, which, however rich, is exposed to the interpretative error of automatism.

From the conceptual point of view, the vehicular representation used in one or more embodiments can be thought of as a union of points, spatially connected in a reticular structure of appropriate density, equipped with 3D coordinates and structure (texture - eg. color) which defines its visible attributes.

One or more embodiments can make use of numerical methodologies in a system with cameras SC1, SC2 metrologically calibrated both from a geometric and a radiometric and chromatic point of view, as discussed below, capable of providing 3D reconstructions with established accuracy.

One or more embodiments can implement a precise spatial-temporal positioning of the vehicle in the points close to the detection stations (eg. A and B) with the ability to control the speed error, especially on short road sections.

One or more embodiments can use, in addition to the stereoscopic setting, which is richer in information than the single camera, the availability, current time, of a very high number of elementary cells (pixels), together with the relative sensitivity and the dynamic camera sensors.

One or more "simple" sensors, such as a single camera, can provide an incomplete and non-invariant 2D representation of the vehicle 3D object, in the sense that this representation can change unpredictably as it passes from point A to point B.

This makes it difficult to establish a direct identity at the raw data level between the two observations of the same vehicle W.

The use of a non-invariant signal may require a subsequent process to be carried out in order to obtain a description at a higher level of abstraction, precisely invariant, at least of the most discriminating parts of the vehicle: this can happen, for example. through automatic recognition methods, for example the OCR reading of textual parts or the recognition of lines or graphic symbols. However, a process of recognition of these parts can give rise to errors that penalize a priori the association function of the description of the vehicle in transit from A to B. Furthermore, it is not generally possible to establish a priori which of these attributes are actually more discriminants in a generic application scenario.

In one or more embodiments, the complete representation of the vehicle obtained with "complex" sensors such as e.g. the stereo-cameras SC1, SC2 allows to provide a complete and invariant 3D representation of the 3D object, that is, of the vehicle.

An invariant signal allows a direct association, that is not mediated by processes of recognition, potentially fallible, of the description of the vehicle in transit.

More explicitly, the 3D surfaces detected at the entry / exit points, A and B, “colored” by their respective textures, can simply be rigidly rototranslated until they overlap each other.

This process can be facilitated by the analysis of the 3D motion that can be made available on both devices SC1-U1, SC2-U2, with the ability to provide, virtually in real time, information on the relative alignment of surfaces, eg. based on the observed directions of motion, and on the presence of the road as a motion constraint.

In one or more embodiments, this can drastically limit the geometrically admissible transformations, making sure that the evaluation of identity or not no longer depends on "anthropic" concepts such as text, symbols, etc. but only from the superimposition of the colored surfaces, which can be evaluated numerically, and in a much more reliable way, without any reference to such concepts.

In particular, it is unnecessary to carry out a (correct) OCR reading of the license plate for the evaluation of correspondence, this reading only facilitating access to the data of the vehicle holder by the system operators.

The fact of being able to avoid the use of an OCR reading has the advantage of freeing oneself for example. from the complexity of the plates of the single nations, conceiving a system intrinsically applicable all over the world. This aspect also allows the extension to similar systems but not oriented to road vehicle traffic control applications, for example industrial areas, airport, port, marine areas where traffic involves vehicles not necessarily equipped with license plates and more generally it allows the use in cases in which the identity between two objects observed in different conditions is to be established, on the basis of their shape / color.

The representation of the vehicle previously called 3D + Tx for brevity (three-dimensional texture representation) is one of the most conservative and reliable obtainable from optical sensors (cameras) and facilitates a priori a high robustness of the association process between A and B.

In one or more embodiments, the association can take place successfully even in the event of a partial lack of correspondence, for example in the case of partial occlusions of vehicles (due to dense traffic) or in the presence of local disturbances (luminous reflections, headlights on ).

In fact, the association process between the two 3D + Tx representations does not require the single representations to coincide in every single point. By virtue of the quantity of identifying information of the vehicle and distributed on its surface, any partial occlusion of the 3D + Tx representation caused in A and / or in B of the presence, for example of another vehicle, still allows the successful execution of the process of association between A and B.

Furthermore, one of the two 3D + Tx representations of the same vehicle taken in and out may not partially or totally include the area of the surface relating to the plate, without affecting the ability of the system to perform the correct association. It is sufficient that only the visible part in both points A and B contains 3D + Tx information such as to make the association unambiguous.

All this in contrast with the symbolic and abstract description of a very small part of the vehicle which is the license plate, which could make the association bankrupt in case of lack of visibility of the plate itself.

The 3D + Tx representation can also provide a direct response to the application requirement of the variable speed limit by vehicle class (e.g. light vehicle and heavy vehicle), which can be determined directly by means of a specific automatic classification system based on the observed size . In this case, the use of the plate alone would make it impossible to obtain this type of information.

From a technological point of view, the 3D reconstruction and precise positioning of the vehicle with respect to the camera can be achieved through geometric calibration of the stereometric structure, which is maintained and kept under control during operation: the structural solutions presented EP 1 997 090 A1 (already mentioned - see for example Figure 7), make it possible to achieve this goal.

Methodological details on the methods of joint estimation of motion and 3D structure of objects can be found in the scientific literature: see in this regard, eg. G.P.Stein, A.Shashua: "Direct Estimation of Motion and Extended Scene Structure from a Moving Stereo Rig", MIT Technical Report, 1997; N.Molton, M. Brady "Practical Structure and Motion from Stereo When Motion is Unconstrained" International Journal of Computer Vision, Volume 39, Issue 1, August 2000, Pages 5-23; L. Valgaerts et al. "Joint Estimation of Motion, Structure and Geometry from Stereo Sequences", Proceedings of the 11th European Conference on Computer Vision: Part IV, 2010, Pages 568-581.

In one or more embodiments, as exemplified in Figure 2, a processing unit U1 configured to perform the following functions can be associated with the (first) stereo-camera SC1:

- 100a: stereo-camera image acquisition and time signal;

- 101a: functional check (active time synchronism, device calibration, highway device calibration);

- 102a: report generation;

- 103a: parameter control (stereo-camera exposure SC1, internal-external clock);

- 104a: 3D-Tx report generation

- 105a: sending 3D-Tx report to 106 units.

In one or more embodiments, as exemplified in Figure 2, the (second) stereo-camera SC2 can be associated with a processing unit U2 essentially similar to the U1 unit and therefore configured to perform the following functions:

- 100b: stereo-camera image acquisition and time signal;

- 101b: functional check (active time synchronism, device calibration, highway device calibration);

- 102b: report generation;

- 103b: parameter control (stereo-camera exposure SC1, internal-external clock);

- 104b: 3D-Tx report generation

- 105b: sending 3D-Tx report to 106 units.

In one or more embodiments, wishing to verify the identity of the vehicle also from the point of view of the texture (Tx), it is possible to proceed with a radiometric and chromatic calibration of the images (steps 103a-103b), such as to make the data comparable. observed in different lighting conditions.

Furthermore, in one or more embodiments, the fact of generating a 3D + Tx surface through stereometry, integrated by the motion analysis (steps 104a-104b) can take on the character of a numerical process of the black box type, which does not require the specification or learning of symbolic details. It can operate directly on the pixels of the images in a "blind" way, based only on the calibration information of the cameras and on the optical-geometric constraints of the three-dimensional vision.

One or more embodiments can therefore differ substantially from OCR or pattern recognition methods, which provide some kind of description, of a predefined type or learned through statistics, of what is to be detected in the images.

In view of night use, each SC1-U1 and SC2-U2 device can be equipped with illuminators operating in the spectrum of visible light or infrared (not visible in the figures), while during daytime use, when a supplement of light artificial may not be necessary, it is reasonable to assume that the events in A and B occur in a limited period of time and at a limited spatial distance and such that the variations (e.g. sunlight alternately masked by clouds) are irrelevant and / or so small that they can be managed through the automatic modification of the acquisition parameters of each SC1, SC2 stereo camera.

In order to have synchronized devices (see steps 100a, 101a and 100b, 101b) in one or more embodiments it is possible to provide that the absolute time is transmitted via the TCP / IP network, capable of guaranteeing transmission everywhere, even in tunnels or others in places without satellite time signal coverage, starting from a universal clock to the two devices SC1-U1, SC2-U2. In one or more embodiments, instant transmission delays and errors can be checked using two further local counters which, although less accurate than a "universal" clock, are not affected by transmission disturbances.

With regard to the observability of the system / process, in one or more embodiments, the single device SC1-U1, SC2-U2 based on stereo-cameras can be designed to constantly check (steps 103a-103b) its own set-up, i.e. calibration, of the intrinsic and extrinsic geometric and optical parameters that allow to extrapolate from the 2D stereo images the three-dimensional information of the vehicle and its position with respect to the real world, that is to point A or B of the road.

Furthermore, in one or more embodiments, each of the devices can provide the observability of this calibration, and this translates into application terms in the possibility of providing evidence of the correct behavior of the system and therefore of the reliability of the measurement of average speed V.

By focusing attention on Figure 1, it will be noted that, in one or more embodiments, both the first SC1-U1 system and the second SC2-U2 system are able to "capture" the vehicles in transit respectively in A in entry to the section and at B exit, with the ability to reconstruct the shape / texture using 3D computer vision algorithms (per se known). With similar methodologies, the 3D motion, i.e. the vector of the input speed, can be determined from their observation in several successive instants (see e.g. tA1, tA, tA2; tB1, tB, tB2 in Figure 1) to estimate them through simple procedures of interpolation the crossing time of the two virtual lines at points A and B at the beginning and end of the section.

As used here, the term texture (structure) identifies the set of points, lines and shapes that can give a surface a compact and unitary appearance. In the graphics typical of 3D representations, textures play a role similar to that of a dress that covers the polygons of the 3D representation themselves, covering them in order to better define them, color them and ultimately make them more realistic to the eye, improving the overall quality of the graphic aspect of a 3D representation; in the specific case of a search for correspondence between A and B, the texture can provide identification information complementary to 3D and equally discriminating in the automatic association process.

For simplicity, as far as it is of interest for the purposes of this exemplary description, it is possible to think of the texture data detectable by the SC1 and SC2 cameras as data inherent to the vehicles W and capable of including, in addition to the overall color / s of the bodywork, also data of greater detail such as the presence and distribution of bands and elements of different colors, etc… which can be used to integrate the 3D representation of the vehicle.

More technically, the texture represents the content of one or more 2D optical images that are superimposed by geometric deformation on a 3D model expressed in the form of a digital surface (e.g. polygonal model), in order to add pictorial and not just morphological details. . In computer graphics, the texture therefore significantly increases the realism of the representation and also allows the reproduction of lighting effects. As an example of 3D + Tx representation techniques we can think of the “wrapping” function of the terrain relief and 3D buildings in satellite maps such as Google Earth®.

In the present context, the role of such a representation may be to match the geometric information present in the 3D model with the pictorial information (colors, edges, etc.) present in the 2D images.

In one or more embodiments, the system can still annotate (record) - including them in the reports sent (steps 105a-105b) to the association procedure (106 - discussed below) also further synthetic data such as e.g. the plate, the model, the manufacturer, the overall color, the presence of coded trademarks and anything else that is useful to know for subsequent analysis.

In one or more embodiments, the data collected by the first device SC1-U1 can be intended to be processed jointly with the data collected by the second device SC2-U2. This can take place (in whole or in part) in a step 106 which can be implemented in a third unit U3, e.g. according to the scheme exemplified in Figure 2.

In one or more embodiments, without prejudice to the location of chambers SC1 and SC2 at points A and B, the "territorial" distribution of units U1, U2 and U3 which are responsible for the various signal processing functions of chambers SC1 , SC2 can be - at least in principle - any, with such units capable of communicating with each other on channels of various kinds, eg. via appropriately protected LAN / WAN network.

In one or more embodiments, the second SC2-U2 device can perform a function essentially similar to the first SC1-U1 device as regards the 3D data / Texture capture function (e.g. color, in the simplified example considered here for illustrative purposes).

In one or more embodiments the devices SC1 and SC2, used in a section traveled in both directions, can perform both the U1 function (input processing) and the U2 function (output processing) depending on the estimated travel direction based on the velocity vector.

In one or more embodiments, for each extracted vehicle W (described in terms of 3D + Tx) the second device SC2-U2 can activate a comparison procedure (matching), which can be carried out precisely in step 106 together with the calculation of the speed.

All in order to give rise to a survey result, capable of being made available to a user, exemplified here by a DB database, with the elementary "record" comprising:

- the identification of a certain vehicle W that has traveled the stretch or section of length L between A and B, e

- the indication of the (average) speed V with which the vehicle in question has traveled this stretch.

With regard to the comparison procedure, a conventional "one-to-many" association search approach could be prone to error. This approach provides, given an exit vehicle detected by the second device, to select, from among all the incoming vehicles communicated by the first device, only those entered in a time interval defined on the basis of a minimum and maximum duration of the crossing and activate a correspondence search; research designed to evaluate if, among all the vehicles entered in this interval, there is at least one compatible with the current one, and in multiple case select the one that maximizes a similarity score or "score". All this while also deleting from the incoming list all the vehicles that are no longer of interest for future verification procedures. This way of proceeding is exposed to the risk of providing results strongly dependent on the order in which the vehicles are analyzed and of easily leading to association errors if the input list contains pairs of vehicles that are very similar to each other. This eventuality is all the more probable the more the traffic is high and the longer the section to be checked. By way of example, the travel time of a 10Km section with a maximum speed of 100Km / h with a speed range from 50km / h to 300km / h corresponds to a maximum of 12 minutes and a minimum of 2 minutes, therefore the interval of search is 10 minutes. In this interval, with a flow of, for example, 3000 vehicles / hour, an average of 500 vehicles can be contained, with a high probability of finding very similar ones.

One or more embodiments, on the other hand, can adopt a more robust matching function, by matching not between a single outgoing vehicle and all incoming vehicles that are temporally compatible, but between a list of outgoing vehicles, collected in a certain window. time, and all the incoming vehicles contained in a search interval based on similar criteria based on the minimum and maximum crossing time of the section.

Such a type of association, called "many-to-many" to distinguish it from the "one-to-many" association discussed above, allows to maximize the reliability of the matching even in the case of the presence of similar vehicles at the entrance / exit at short notice, the correct management of which could be jeopardized in the case of the simple "one-to-many" process.

This type of association, of a completely general type, is described in the scientific literature: see eg. J. Munkres: "Algorithms for the Assignment and Transportation Problems", Journal of the Society for Industrial and Applied Mathematics, Volume 5 Issue 1, March 1957, Pages 32–38.

This association can only be based on numerical data (matrix of similarity measures) and therefore does not require further complex analysis at the level of images.

In other words, in one or more embodiments, the "similarity" of two generic vehicles X and Y can be established by placing points of the two surfaces 3D + Tx in direct correspondence, using the reference position and the velocity vector as data to get the best possible alignment.

Aligned the two surfaces, in one or more embodiments it is possible to evaluate the degree of correspondence, for example. through a (weighted) sum of the locally detected differences, both of geometric type and of structure (texture, for example color). In this way it is possible to establish an objective indicator (score or "score") of similarity that can include both dimensional and pictorial aspects, and regardless of any interpretation of such data aimed at extracting a semantic meaning.

In one or more embodiments, using the principle of "many-to-many" correspondence, the search for optimal associations can initially create a matrix of similarity measures between all the possible pairs between incoming X vehicles (point A) and Y vehicles at the exit (point B), as schematically represented in Figure 3 where it is highlighted e.g. the possible correspondence between Win vehicles (..., Xn-2, Xn-1, Xn, Xn + 1, Xn + 2, Xn + 3, Xn + 4, Xn + 5, ...) detected in sequence at the entrance to the section subject to detection in a certain time window on entry and Wout vehicles (..., Yk-2, Yk-1, Yk, Yk + 1, Yk + 2, Yk + 3, Yk + 4, ...) sequentially detected on exit in a window outgoing time compatible with the incoming window. By "compatible" we mean an exit time window delayed with respect to the entry time window of a time interval related to the possible travel times of the section of length L by vehicles W (taking into account the possible ranges of speed variation that you want to be able to detect).

On the basis of this matrix it is possible to carry out an association procedure with the aim of identifying the set of couples that is globally better; in the presence of ambiguous solutions (eg vehicles that are completely identical or in any case non-discriminable based on the available data), it is also possible to report the anomaly of a possible double association. Once the best possible matches have been established, it is finally possible to evaluate the section speeds and, if necessary, to generate the appropriate signals.

For example, referring to the matrix representation (purely exemplary) of Figure 3 it is possible to see that the vehicles identified (respectively in input and output) as Xn-2-Yk, Xn-Yk + 2, Xn + 2-Yk + 3 , Xn + 5-Yk + 4 appear to have kept their relative positions, an indication that they have traveled the section more or less at the same speed.

The vehicles identified (always respectively at the entrance and at the exit) as Xn + 1-Yk-1 and Xn + 3-Yk + 1, demonstrate instead that they have traveled the section faster than others, overtaking them.

The vehicle identified (once again respectively at the entrance and exit) as Xn + 4-Yk-2 shows instead of having traveled the section at high speed, having in practice overtaken all the vehicles that the 3D interpolation of the trajectories of the vehicles. The crossing time of the section can be determined on the basis of the difference in the crossing times, which is possible thanks to the synchronization between the two devices SC1-U1, SC2-U2, achievable in the ways mentioned above.

In one or more embodiments, the information provided to the user DB may also include further information in addition to the identification of a certain vehicle W and the indication of its speed.

The nature of these additional indications (e.g. number plate, overall color of the vehicle, type, model) and the same methods of presentation of the information (image, text, etc ...) can be chosen differently depending on the needs or the nature of the user (road patrol, surveillance center, municipal administration, etc.), which can be the most diverse.

Returning to the topic of distribution on the territory of units U1, U2 and U3 of Figure 2, in one or more embodiments, the correspondence search function (unit U3) can in fact be "hosted" on the second device (unit U2) , for example assuming that the reconstruction of the output shape / color attributes can use a software asymmetry, in the calculation of the 3D reconstruction, to maintain the hardware symmetry of the two devices.

In one or more embodiments, the use of a representation of the 3D + Tx data (3D and texture) in multi-scale format can facilitate the choice of a reasonable compromise between accuracy of the description and computational complexity. Such a representation can therefore simultaneously host all possible investigation scales, corresponding to different levels of detail, through a hierarchical data structure known in the IT world as octree.

This type of representation allows you to efficiently describe volumes and surfaces in 3D and is commonly used in computer graphics. More details can be found in the scientific literature, eg. in C.H.Chien, J.K.Aggarwal: "Volume / surface octrees for the representation of threedimensional objects", Journal Computer Vision, Graphics and Image Processing, Volume 36 Issue 1, Oct.1986, Pages 100-113.

In one or more embodiments, the similarity measurement can therefore be limited only to the coarser scales in the case of vehicles that are immediately reliable on the basis of the length L of the (known) section and of the 3D positions in the instants of detection, i.e. Tin and Tout, e

- the time T = Tout - Tin elapsed between these instants. In one or more embodiments, the measurement of the elapsed time may involve a procedure (steps 103a, 103b) for correcting any errors of a systematic type. In the event that errors beyond a predetermined limit are detected, the procedure can produce an anomaly signal and interrupt the speed calculation until the correct synchronization is restored.

In formal terms, indicated with EV the percentage error on speed and called L the length of the section, V the speed and T the travel time, is

It is possible to express this error EV through the relationship:

∆ V ∆ L ∆ T

Ε V = =

V L T

Observing that L = V ∙ T, the same relation can be reformulated as

∆L + V ∆ T

Ε V =

L

The latter expression highlights the fact that the speed error is an increasing function of the speed V and a decreasing function of the length of the section L. Its maximum value is therefore obtained for a section of minimum length traveled at the maximum detectable speed. In other words, the above relationship allows you to evaluate the maximum speed below which the error EV still remains within an assigned value, known the maximum values of ∆L and ΔT.

A system with metrological characteristics as exemplified here facilitates compliance with these values.

Analyzing separately the sources of error ∆L and ΔT, it is observed that they are affected by both systematic and random errors.

In the case of the temporal component ΔT, the instrumental part can be considered negligible, while the random one can be managed through synchronization, foreseeing for example. that in the absence of it, the speed measurement is not generated.

In the case of the spatial component ∆L, the systematic part is linked to the knowledge of the length L of the section (obtainable for example during the installation and calibration of the system) while the random part depends on the ability of the system to correctly position the vehicle W with respect to at the entry / exit points and to be able to demonstrate this correctness.

It has been observed that one or more embodiments allow to overcome the limitations of various known solutions also at the level of certification of compliance with the accuracy requirements imposed.

In fact, having only the 2D images it is not possible to establish in a sufficiently certain way the distance between the vehicle and the shooting point (or other reference point): small variations in the aiming of the camera or in its optical parameters are invisible to observation but can induce variations. not negligible spatial (eg even several meters); this makes the speed measurement valid only for sections long enough to make the weight of this type of error negligible.

Purely by way of example - a system based on 2D cameras with a license plate reading process and therefore devoid of precise and verifiable 3D references, may have a positioning error of the vehicle in A and B of a few meters.

This leads to exceeding the maximum reference value of 1% error already on 500m sections. If we add to this the error component due to the synchronization (equal to 0.1 seconds with the technology currently available) and to the speed of a hypothesized offending vehicle equal to 150Km / h, the overall error exceeds the value of 1% even on longer routes.

One or more embodiments, including (at least) two 3D + Tx capture devices and a correspondence search function and consequent speed calculation, can therefore use a 3D representation of the vehicle (e.g. in Cartesian coordinates common with the camera ) such as to allow the detection (with at most an error of a few centimeters) the relative position of the observer vehicle, thus effectively eliminating the impact of the random component ∆L, making it significantly lower than the systematic component.

A 3D description is in fact more complete than a 2D description of the same object made through a list, however long, of attributes. This list, which requires individual recognition processes, in addition to not being able to be exhaustive, inevitably accumulates errors. In addition to being able to provide greater accuracy than the representation of the vehicle itself, the 3D information allows an accurate measurement of the position with respect to points A and B, known a priori the 3D positions of the stereo cameras SC1, SC2 with respect to the points themselves. This makes possible the metrological verification of the accuracy of the distance of the vehicle in A with the vehicle itself in B and then, using the elapsed time, the metrological verification of the accuracy of the physical quantity of application interest, ie the section speed.

A purely 2D description, obtained from a single camera, does not allow to identify, if not roughly and without certain metrological references, the position of the vehicle with respect to the two crossing lines of the road at points A and B and therefore is not able to set limits stringent to the possible numerical error in the speed estimate.

In one or more embodiments, such an error can instead be completely controlled both from the spatial point of view (position of the vehicle with respect to points A and B) and from the temporal point of view (synchronization control).

In this regard, it can be observed that the request for certifiable precision is not only a technical fact but has a considerable importance for possible sanctioning purposes: a high error effectively reduces the number of detectable violations, given the need to maintain a wider tolerance interval. and the fact that most violations occur for speeds slightly above the maximum tolerated limit. Such a system is in fact exposed to the risk of having to suspend the detection in case of lack of spatial (camera moved) or temporal (loss of synchronization) feedback.

It has also been observed that various known solutions provide for rear reading only (obtained by framing the vehicle from the rear), excluding the possibility of front reading (obtained by framing the vehicle from the front). This presumably in consideration of the fact that the front reading can be more difficult due to the reduced size of the front plates, if not the lack of provision of the same in vehicles circulating in certain states.

One or more embodiments can completely disregard the type of installation, which can therefore be both rear (as schematically represented in Figure 1), and front, or - at least in principle - lateral (i.e. with an additional camera that frame the back for documentation purposes only), as there is no need to read the license plate in real time.

One or more embodiments therefore overcome the intrinsic limitations of the systems that explicitly reduce the entire vehicle to the plate only, completely ignoring any other type of information useful both for the evaluation of correspondence and for the estimation of speed.

Moreover, using a plate reading system operating continuously, the reading can take place at any point in the observation cone where the plate can be detected by the camera that frames it. In such conditions, the plate ends up not referring to a precise spatial position, even when proceeding - eg. through 2D interpolations between two frames - to somehow place the license plate in a 3D space.

One or more embodiments may be based, e.g. on an interpolation between subsequent 3D surveys ("frames") and calculate with minimal uncertainty the instant in which a vehicle W crosses the virtual lines of the start (point A) and end (point B) of the section of the route being surveyed.

However, one or more embodiments can completely disregard the plate and the syntactic correctness of its reading, based instead on the invariant geometric and pictorial characteristics of the vehicle itself, i.e. on its geometric shape and its visual appearance determined by the texture.

It will be appreciated that one or more embodiments may however provide for the detection of the plate which (as a highly identifying pictorial datum) can be used as a further element of comparison. All in conditions in which the non-automatic reading does not invalidate either the definition of the input-output correspondence nor the estimate of the section speed.

In one or more embodiments, a method for detecting the speed (e.g. the average speed V) of vehicles (e.g. W) in a route section having a length L between an entry point (e.g. A ) and an exit point (e.g. B) can include:

- detecting (e.g. 100a-104a) the transit of said vehicles in said entry point with first visual detector means (e.g. SC1) and in said exit point with second visual detector means (e.g. SC2), obtaining (e.g. in U1, U2) from said first and second visual detection means respective representations (e.g. 105a, 105b) of the vehicles in transit,

- processing (e.g. U3, 106) the respective representations obtained by said first and second visual detection means by coupling (i.e. associating e.g. by identifying) a certain vehicle (e.g. ..., Yk-2, Yk-1, Yk, Yk + 1, Yk + 2, Yk + 3, Yk + 4, ...) in transit at said exit point at an exit time (e.g. Tout) with a vehicle (e.g. ..., Xn-2, Xn -1, Xn, Xn + 1, Xn + 2, Xn + 3, Xn + 4, Xn + 5, ...) in transit in said entry point at an entry time (eg.

Tin), calculating the speed of said certain vehicle on said stretch of route as a function of said length L and of the time elapsed between said entry time and said exit time (e.g. V = L / T where T = Tout - Tin ).

In one or more embodiments, the method can comprise said respective representations of the vehicles in transit as three-dimensional (3D) representations of the vehicles themselves.

In one or more embodiments, the method can comprise said first visual detector means and said second visual detector means as stereometric detector means (stereo-cameras).

One or more embodiments can provide for configuring said first visual detecting means and said second visual detecting means to obtain said respective three-dimensional representations of the vehicles themselves by extrapolating two-dimensional representations of the vehicles.

One or more embodiments can provide for obtaining said respective representations of the vehicles in transit as three-dimensional representations of the vehicles themselves supplemented by structure information (texture).

In one or more embodiments, said structure information (texture) can comprise information on the color of said vehicles.

In one or more embodiments, said processing (U3, 106) the respective representations obtained by said first and second visual detecting means can comprise a matching procedure in which the correspondence between:

- a set of representations obtained by said second visual detection means for a plurality of vehicles (e.g. Yk-2, Yk-1, Yk, Yk + 1, Yk + 2, Yk + 3, Yk + 4, ... in Figure 3) in transit at said exit point, e

- a set of representations obtained by said first visual detection means for a plurality of vehicles (e.g. ..., Xn-2, Xn-1, Xn, Xn + 1, Xn + 2, Xn + 3, Xn + 4, Xn + 5,… in Figure 3) in transit at said entry point.

One or more embodiments can provide for evaluating the correspondence between the representations of said sets by means of a sum, preferably weighted, of the differences between said representations.

One or more embodiments can provide for: - obtaining from said first and second visual detection means representations of the vehicles in transit having at least a first level and at least a second level, respectively coarser and finer,

- evaluating the correspondence between the representations of said sets starting from said first level, extending the evaluation of the correspondence to said at least a second level only for the representations for which the correspondence to said at least one first level has been found.

In one or more embodiments, a system for detecting the speed (e.g. average) of vehicles in a section of the route having a length between an entry point and an exit point, can comprise:

- first stereometric visual detector means which can be positioned at said entry point,

- second stereometric visual detector means which can be positioned at said exit point,

- processing means (e.g. U1, U2, U3) which can be coupled to said first and second visual detection means to obtain respective representations of the vehicles in transit,

the system being configured to operate with the previously exemplified method.

Finally, it will be appreciated that the possibility of obtaining representations of vehicles in transit as three-dimensional representations of the vehicles themselves is not limited to the solution exemplified here, based on the use of stereometric detection means such as stereo cameras, operating for example. with techniques of extrapolation of two-dimensional representations of vehicles.

In one or more embodiments, to obtain representations of the vehicles in transit as three-dimensional representations of the vehicles themselves, it is possible to resort to different solutions such as e.g. time-of-flight cameras (Time Of Flight camera or TOF-camera) or profilometry techniques.

A time-of-flight camera allows you to evaluate the distance between the camera and an object - eg. a vehicle - detecting the "flight" time taken by a light pulse to travel the round trip between the camera and the object.

The detection of the distance can be carried out selectively (ideally for each pixel), which makes it possible to obtain the respective representations of the vehicles in transit as three-dimensional representations of the vehicles themselves.

In one or more embodiments, whatever the solution adopted to obtain three-dimensional representations of vehicles in transit, an optical camera can provide "pictorial" information for the 3D + Tx representation.

Without prejudice to the basic principles, the construction details and the embodiments may vary, even significantly, with respect to what is illustrated here purely by way of non-limiting example, without thereby departing from the scope of protection.

This scope of protection is defined by the attached claims.

Claims (10)

CLAIMS Method for detecting the speed (V) of vehicles (W) in a section of the route having a length (L) between an entry point (A) and an exit point (B), the method comprising: - detecting (100a-104a) the transit of said vehicles (W) in said entry point (A) with first visual detecting means (SC1) and in said exit point (B) with second visual detecting means (SC2), obtaining (U1, U2) from said first (SC1) and second (SC2) visual detecting means respective representations (105a, 105b) of the vehicles (W) in transit, - elaborate (U3, 106) the respective representations (105a, 105b) obtained from said first (SC1) and second (SC2) visual detection means by coupling a certain vehicle (..., Yk-2, Yk-1, Yk, Yk + 1 , Yk + 2, Yk + 3, Yk + 4, ...) in transit at said exit point (B) at an exit time with a vehicle (..., Xn-2, Xn-1, Xn, Xn + 1, Xn + 2, Xn + 3, Xn + 4, Xn + 5, ...) in transit in said entry point (A) at an entry time, - calculating the speed of said certain vehicle on said stretch of the route as a function of said length (L) and of the time elapsed between said entry time and said exit time, the method comprising obtaining said respective representations (105a, 105b) of the vehicles (W) in transit as three-dimensional representations of the vehicles themselves. A method according to claim 1, comprising configuring said first visual detector means (SC1) and said second visual detector means (SC2) as stereometric detector means. Method according to claim 2, comprising configuring said first visual detection means (SC1) and said second visual detection means (SC2) to obtain said respective three-dimensional representations of the vehicles themselves by extrapolating two-dimensional representations of the vehicles (W). Method according to any one of the preceding claims, comprising obtaining said respective representations (105a, 105b) of the vehicles (W) in transit as three-dimensional representations of the vehicles themselves supplemented by structure information (texture). Method according to claim 4, wherein said structure information (texture) comprises information on the color of said vehicles (W). Process according to any one of the preceding claims, wherein said processing (U3, 106) the respective representations (105a, 105b) obtained from said first (SC1) and second (SC2) visual detecting means comprises a matching procedure in which one evaluate the correspondence between: - a set of representations obtained from said second (SC2) visual detection means for a plurality of vehicles (..., Yk-2, Yk-1, Yk, Yk + 1, Yk + 2, Yk + 3, Yk + 4, ... ) in transit at said exit point (B), e - a set of representations obtained from said first (SC1) visual detection means for a plurality of vehicles (..., Xn-2, Xn-1, Xn, Xn + 1, Xn + 2, Xn + 3, Xn + 4, Xn + 5, ...) in transit at said entry point. 7. A method according to claim 6, comprising evaluating the correspondence between the representations of said sets by means of a sum, preferably weighted, of the differences between said representations. 8. Process according to claim 6 or claim 7, comprising - obtain from said first (SC1) and second (SC2) visual detection means representations (105a, 105b) of the vehicles (W) in transit having at least a first level and at least a second level, respectively coarser and finer, - evaluate the correspondence between the representations of said sets starting from said at least one first level, extending the evaluation of the correspondence to said at least a second level only for the representations for which the correspondence to said at least one first level has been found. 9. System for detecting the speed (V) of vehicles (W) in a section of the route having a length (L) between an entry point (A) and an exit point (B), the system comprising: - first stereometric visual detector means (SC1) which can be positioned in said entry point (A), - second stereometric visual detector means (SC2) which can be positioned in said exit point (B), - processing means (U1, U2, U3) which can be coupled to said first (SC1) and said second (SC2) visual detection means to obtain respective representations (105a, 105b) of the vehicles (W) in transit, the system being configured to operate with the method according to any one of claims 1 to 8. 10. Computer product that can be loaded into the memory of at least one processing device (U1, U2, U3) and comprising portions of software code for carrying out the steps of the method according to any one of claims 1 to 8 when the product is executed on at least one processing device processing.