JP6236687B2 - Overhead inspection system using time series filter processing - Google Patents

Description

  The present invention relates to an overhead wire inspection device for inspecting and measuring the displacement of overhead wires using a time series filter.

  In Non-Patent Document 1 (Genshiro Kitagawa, “Monte Carlo Filter and Smoothing,” Mathematical Statistics, Vol. 44, No. 1, pp. 31-48, 1996), the Monte Carlo filter (particle filter) is a time series filter. This is a method similar to the method called “.” (Hereinafter referred to as “particle filter”).

Here, the particle filter will be described with reference to FIG.
First, the upper part of FIG. 1 is a probability density distribution in which the horizontal axis indicates the state and the vertical axis indicates the probability. In this method, the state changes depending on the situation (overlapping line deflection, wear start position, wear width x number of overhead lines), but all of them are multi-dimensional. (For example, it can be regarded as a function representing the probability of overhead line deviation). The upper left of FIG. 1 shows a uniform probability density distribution of states (overhead position, wear width, etc.), and the upper center of FIG. 1 calculates the state distribution by applying a state transition model to the posterior distribution.

Next, the middle part of FIG. 1 shows the probability density distribution displayed in a discrete manner. A continuous function such as a probability density distribution cannot be calculated when it cannot be expressed by a mathematical expression, but a value close to the original probability density distribution can be calculated by discretization and approximation. This is shown as “approximation of probability density distribution” in the figure.
Here, it is discretized using the Monte Carlo method (a method of approximating with a product sum of discretized values instead of obtaining an integral value of a continuous function). In short, the discretization is performed densely at locations showing a high value in the probability density distribution and sparsely discretized at low locations.

In the middle left of Fig. 1, the probability density distribution is uniform, so discretization is performed at regular intervals. Further, the points expressed when discretized are called particles.
In the lower part of FIG. 1, whether or not the discretized particles obtained in the middle part of FIG. 1 are likely is evaluated. For this, observation data (for example, data in which information on overhead line deviation is observed) is used. Here, a high value is given if the particle is plausible, and a low value is given if it is not plausible. This is called likelihood. Likelihood is calculated from the observed data (posterior distribution).

A posteriori probability (posterior distribution) is obtained by performing likelihood evaluation on all particles. This is because if the probability density distribution given first is the prior probability, the posterior probability is obtained by Bayes' theorem using likelihood. Therefore, the lower part of FIG. 1 represents the posterior probability.
FIG. 1 shows how this changes according to the observation data that changes with time, as indicated by arrows from left to middle to right. Actually, it does not end on the right, but this process is performed for the time. This is described as “repeat” in the figure.

  In addition, “the state distribution model is calculated by applying the state transition model to the posterior distribution in the upper center of FIG. 1”. Here, the “state transition model” means, for example, an overhead line displacement is If it is the position of x1, it is the probability of becoming the position of x2 at the next time, and by using a state transition model (also referred to as state transition probability, state transition equation) for the posterior distribution of the previous time, A new (current time) prior distribution can be obtained.

  Incidentally, the prior distribution actually obtained is discretized. In FIG. 1, the lower left posterior distribution is used in terms of processing, and the middle middle prior distribution is obtained. The upper part of FIG. 1 is an image diagram showing what happens if this prior distribution is represented by a continuous function (probability density distribution).

  In Patent Document 1 (Japanese Patent Application Laid-Open No. 2012-208095, “Trolley Wire Measuring Method and Device”), a trolley wire measurement is performed in which light is projected and the reflected light is received to measure the outer shape of the trolley wire (overhead wire). In the method, an image of the trolley line and the nearby train line equipment is taken in the rigid train line section, the deviation of the train line equipment is measured based on the photographed image, and the measurement result is used to measure the outer shape of the trolley line. It is a technique to reflect.

  In Patent Document 2 (Japanese Patent Application Laid-Open No. 2008-298733, “Trolley Wire Wear Measuring Device by Image Processing”), images of several lines obtained by a line sensor are processed together to measure temporal wear of overhead wires. Can be done.

  In Patent Document 3 (Japanese Patent Application Laid-Open No. 2009-274508, “Trolley Wire Wear Measurement Method and Measurement System”), noise is removed from an image obtained by a line sensor by using an average value of a background portion other than an overhead line. In addition, the wear of the overhead wire is measured by using the kilometer information.

JP2012-208095 JP2008-298733 JP2009-274508

Genshiro Kitagawa, "Monte Carlo Filter and Smoothing," Mathematical Statistics, Vol. 44, No. 1, pp. 31-48, 1996

  The method of Non-Patent Document 1 is a paper on particle filters related to the present invention. In the present invention, this particle filter is applied to perform overhead wire inspection.

  The method of Patent Document 1 is a method of measuring a new overhead line using the already-measured deviation of the train line equipment and overhead line, but it is a deterministic method, and the result is added to the existing measurement data. There is a possibility to be heavily dependent.

  In the method of Patent Document 2, the overhead wires obtained by the line sensor are processed together for several lines, and the wear of the overhead wire can be measured with high accuracy by checking the connectivity. It is a problem that the result changes depending on how to separate several lines.

  In the method of Patent Document 3, high-precision overhead wire wear is measured by utilizing not only image processing background removal but also kilometer information, but it is a deterministic technique similar to Patent Documents 1 and 2, The problem is that the result changes depending on the error in kilometer information.

  The overhead line inspection device according to claim 1 of the present invention that solves the above-described problem is an overhead wire inspection device that accurately measures overhead lines on a train by image data processing using a stochastic time series filter. An image input unit for inputting image data including deviation, width, and wear width of the overhead wires taken in time series by a line sensor disposed on the upper surface of the train; and deviation, width, and wear of the overhead wires A likelihood initialization function using a particle initialization unit that prepares particles having random parameters of width, the particles prepared by the particle initialization unit, and the image data of the previous time input by the image input unit A likelihood calculation unit for calculating a likelihood value based on the likelihood, and a resampler for selecting a particle having a high likelihood value with a high probability based on the likelihood value calculated by the likelihood calculation unit A resampler for performing, characterized in that it consists of a state transition unit for creating a new particle is varied based on the parameters of the particles resampled to the state transition model by the resampler.

  The overhead line inspection / measurement apparatus according to claim 2 of the present invention for solving the above-mentioned problems is characterized in that, in claim 1, the overhead lines are single or a plurality of overhead lines.

The overhead line inspection / measurement apparatus according to claim 3 of the present invention that solves the above-described problems is characterized in that, in claim 2, the overhead lines include a suspension line.
The overhead line inspection / measurement apparatus according to claim 4 of the present invention for solving the above-described problems is characterized in that, in claim 2 or 3, the overhead lines include a crossover.

  The overhead line inspection / measurement apparatus according to claim 5 of the present invention for solving the above-mentioned problem is the overhead line inspection / measurement apparatus according to claim 1, wherein the line sensor includes a plurality of line sensors, and also measures the height of the overhead lines, and the initial particle The conversion unit includes the height of the overhead lines in the parameter.

  The overhead line inspection / measurement apparatus according to claim 6 of the present invention for solving the above-described problem is the overhead line inspection / measurement apparatus according to claim 1, wherein the likelihood function includes an absolute value of a difference between pixel values before and after a wear region of the overhead line, and the overhead line The difference value when the average pixel value of the overhead line area is subtracted from the average pixel value of the wear area of the kind, the deviation / width / wear width of the overhead line at the previous time, and the deviation of the overhead line at the current time -Based on the difference between width and wear width.

(1) Overhead lines based on prediction using image information of the previous time, that is, single overhead lines, multiple overhead lines, overhead lines / suspended lines, overhead lines / suspended lines / crossover lines, can be measured.
(2) Previously measured overhead lines, that is, single overhead lines, multiple overhead lines, overhead lines / suspension lines, single overhead lines / suspension lines / transition lines using information on overhead lines / suspension lines / crossover lines as state transition equations Inspection is possible.
(3) Overhead lines using prediction based on probability theory, that is, single overhead lines, multiple overhead lines, overhead lines / suspension lines, overhead lines / suspension lines / crossovers can be measured.
(4) It is possible to increase the accuracy of overhead / suspended / crossover line inspection using multiple cameras and to measure the height of overhead / suspended / crossover lines.

It is a schematic diagram of a particle filter. It is a flowchart of the overhead wire inspection apparatus which concerns on the 1st-4th Example of this invention. It is a block diagram of an overhead wire inspection apparatus according to first to fifth embodiments of the present invention. It is a flowchart which concerns on the 5th Example of this invention. It is a graph which shows the relationship between the previous line and the next line. It is an image figure of the image data imaged by a line sensor in time series.

  There is a desire to accurately measure the trajectory of overhead lines taken from the upper surface of a train using a line sensor camera. Here, in the present invention, the overhead line means a single or a plurality of overhead lines, and further includes a suspended overhead line and a crossover line.

However, it is difficult to detect only the overhead line by simple image processing because various things are measured on the image and there are various factors such as sunlight in the outdoor environment that are difficult to measure.
Therefore, in the present invention, as in Patent Documents 1, 2, and 3, it is considered that light is projected when an overhead line is photographed from the upper surface of the train with a camera.
By projecting light, the wear surface of the overhead line reflects the light, and by measuring it, we aim to realize highly accurate overhead line inspection.

Further, according to the present invention, the characteristics of the overhead line are “the pixel value changes greatly before and after the wear area”, “the average pixel value of the wear area is larger than the average pixel value of the overhead line area excluding the wear area”, “ Using information such as “the overhead wire region and the wear region do not change significantly from the deviation or height of the previous line”, measurement using the image processing of the overhead wire displacement and its wear width is performed.
First, since there is a feature that “the overhead wire region and the wear region do not change significantly from the deviation of the previous line”, the deflection of the overhead wire using a particle filter that is a time series filter effective for the tracking of Non-Patent Document 1 • Estimate wear.

As a feature of the particle filter, prediction using time series information can be performed, and probability theory can be handled instead of determinism.
Since the state transition probability can be introduced into the prediction using the time series information, in addition to the above features, information known by past measurement can be used as a probability model (FIG. 1).

In the particle filter, calculation difficulty is eliminated by discretizing the probability density distribution by sampling.
This sampling point is called a particle. How likely each particle is as an optimal solution is obtained by likelihood calculation.
At this time, how to set the likelihood function is very important.

  In the present invention, the likelihood is calculated by using a likelihood function considering a plurality of elements. First, the state vector of the particle filter is a three-dimensional vector of the displacement of the overhead line in one line imaged by the line sensor, the overhead line width from end to end of the overhead line, and the width of the wear area.

  Next, as a likelihood function, the pixel value greatly changes before and after the wear area, the average pixel value of the wear area becomes larger than the average pixel value of the overhead line area excluding the wear area, the overhead line area and the wear area. A likelihood function that takes into account that the region does not change significantly from the deviation of the previous line is set.

  In other words, the likelihood function includes the absolute value of the difference between the pixel values before and after the wear region, the difference value obtained by subtracting the average pixel value of the overhead wire region from the average pixel value of the wear region, and the deviation of the previous time (previous line). The difference between the current position and the current time (current line) is based on the difference (the higher the distance is within the overhead line width, the higher the value is 0).

Specifically, these three are combined into a likelihood function. At present, the sum of all the values is used, but depending on the case, processing (weighted sum) is performed to increase the weight of each item. The likelihood value is the same as the likelihood.
In addition, since it is known that the shape of the overhead wire does not change significantly, it is possible to estimate a rough displacement of the next line using the previous line information as the displacement information.

By using this information to define a state transition model, high-precision estimation is possible.
Specifically, we currently consider two cases and methods.

  One is a method called “random walk” in which the transition of the overhead line is not known at all. This is a technique for randomly changing the state. This method is often used for particle filters. It is known that the overhead line deviation does not change significantly in the overhead line inspection, so a random walk is performed within the changeable range.

  The other is a case where the shape and self-position of the overhead line are known in advance. In this case, since it is almost known where the deviation of the overhead line will come at the next time, the information is used as a state transition model.

  In the present invention, by using the probability density distribution as shown in FIG. 1, even if there is an error in the estimation on the previous line (the case with the highest probability), an error has occurred in the estimation of the next line. Since information other than the result can also be used, robust estimation of each line is possible as compared with deterministic approaches such as Patent Documents 1, 2, and 3.

  Here, the previous line and the next line mean the data of the previous time and the data of the next time, as shown in FIG. That is, FIG. 5 is a graph in which line data obtained by the line sensor camera is arranged in order of time. In the description of the present invention, the previous line and the next line are, for example, the time for the line data obtained at time t. Data obtained at t−1, data obtained at time t + 1.

In addition, in the present invention, the “detection of overhead lines” → “confirmation of continuity” approach is not used as in Patent Document 2, but “extraction of overhead lines in a continuous area” is performed. The detection accuracy can be improved.
In the description of FIG. 1 described above, the words “Monte Carlo method” and “particle” are used. A time-series filter that processes time-series data using these methods is called a “particle filter” or a “Monte Carlo filter”.

There are various derivation methods such as “important sampling method” and “Markov chain Monte Carlo method” for obtaining the prior distribution from the posterior distribution, but basically all the methods of processing in the flow of FIG. 1 are “particle filter” and “Monte Carlo filter”. Say.
Although the concept is exactly the same, it is a technique often called a particle filter in recent years. Therefore, it is also called a particle filter in the present invention. Since it was called a Monte Carlo filter in the reference, it was explained that it was the same, but it is the same concept. Incidentally, in the present invention, a method called “weighted sampling” is used to obtain a prior distribution.

  There are various types of time series filters, and the “particle filter” used here is one of them. In addition, the “particle filter” is a probabilistic approach, so it can be called a probabilistic time series filter. So we used such names in the claims. Incidentally, other probabilistic time series filters include a “Kalman filter”, but the present invention uses only a particle filter.

Here, “determinism” and “probability theory” can be briefly compared as follows.
Deterministic is to explain the state where the clouds are 30% in the sky as “clear”, and probability theory is to explain that it is “clear with 30% clouds”. The advantage of the probabilistic approach is that not only XX but also the degree can be considered.

“In determinism,“ extract overhead lines ”→“ confirm continuity ”. The problem is that even if you don't know if it's really correct, you run a determinism that says "This part is an overhead line" every time, and later decide that it is correct because it was continuous.
If it is wrong somewhere at one time, or if the overhead line cannot be imaged for some reason, the continuity is interrupted, resulting in erroneous recognition.

On the other hand, the probability theory evaluates the area that seems to be an overhead line at each time and its likelihood.
In the case of the deterministic method described above, a certain point is determined as an overhead line, but in the case of the probabilistic method, the likelihood (of each particle in the present invention) of multiple points is used. Judgment that is not performed can be made.

(1) Basic concept (Example 1)
An object of the present invention is to provide an overhead wire inspection device that accurately measures the displacement, overhead wire width, and wear width of overhead wires by image processing using a time series filter.

  In the following embodiments, as described in Patent Documents 1, 2, 3, etc., as an example, a line sensor (not shown) is installed on the upper surface of the train so as to look up vertically, and light is projected. The luminance signal for one line is photographed by scanning along the direction of the sleeper so as to cross the overhead line, and the contrast between the background and the overhead line in the image in which the luminance signal for the photographed one line is arranged in time series From the difference, as shown in FIG. 6, overhead line displacement, overhead line width, and wear width are acquired in time series by image processing.

  FIG. 6 shows image data photographed in time series by the line sensor at times (t−1), t, and (t + 1). Since the wear width A of the overhead wire 1 is projected, the luminance value is high and becomes a white region, and the overhead wire width B other than the wear width A of the overhead wire 1 has a low luminance value and becomes a black region. The background area is gray between white and black when measured in the daytime. The distance from the reference point C to the wear width A of the overhead wire or the center E of the overhead wire width B is the displacement (direction of sleepers) D of the overhead wire.

FIG. 3 shows an apparatus configuration example according to this embodiment.
As shown in FIG. 3, the overhead line inspection apparatus according to the present embodiment includes an image input unit 10, a storage unit 20, a particle initialization unit 30, a likelihood calculation unit 40, a resampling unit 50, a state transition unit 60, Consists of.

In the image input unit 10, image data for one line captured in time series by the line sensor is input and stored in the storage unit 20. The image data includes time-series overhead wire displacement, overhead wire width, and wear width as shown in FIG.
The storage unit 20 stores various parameter data such as image data, particle data, and state transition probabilities.

In the particle initialization unit 30, random parameters (overhead deflection, overhead line width, wear width) are set in a range and density set for a certain number of particles, and the range, density, and parameters of each particle are stored in the storage unit 20. To do.
The likelihood calculation unit 40 calculates the likelihood using the parameter and image data of each particle based on the likelihood function, and stores the likelihood given to each particle in the storage unit 20.

The resampling unit 50 performs resampling so that particles having a high likelihood are selected with a high probability based on the likelihood. The resampled particles are stored in the storage unit 20, and the particles used for the resampling are deleted.
The state transition unit 60 changes the parameters of the resampled particles based on the state transition model, creates new particles, and stores them in the storage unit 20. According to the new particles that were created, as a probability theory, the overhead wire displacement, overhead wire width, and wear width were measured with high accuracy.

Here, the prior distribution in the initial state is usually unknown. Therefore, as shown on the left of FIG. 1, it is common to assume that the probability density distribution is a uniform distribution.
This uniform distribution is expressed by particles having a random value by the particle initialization unit 30, the likelihood is calculated by the likelihood calculation unit 40 using image data for each particle, and the posterior distribution by the resampling unit 50 Is the procedure of this embodiment.

The “particle filter” is to process time-series data for each time using the Monte Carlo method and particles as described with reference to FIG. Therefore, the flow shown in the flowchart of FIG. 2 becomes the flow of the particle filter itself.
“Image processing using a time-series filter” is a portion in which observation data (image data) is used when likelihood calculation is performed.

A flowchart according to this embodiment is shown in FIG.
First, one line of image data captured in time series by the line sensor is input by the image input unit 10 (step S1).
Next, particles having a random value in the range and density set for each parameter (overhead deflection, overhead line width, wear width) are prepared by the particle initialization unit 30 (step S2).
Subsequently, the likelihood calculation unit 40 calculates the likelihood of each particle using the parameter of each particle and the image data of the previous line based on the likelihood function (step S3).

  Then, a resampling process is performed by the resampling unit 50 based on the likelihood, and a new particle is created by the state transition unit 60 by using the state transition model for the resampled particles (step S4).

  Thereafter, until the time-series imaging by the line sensor is completed (step S5), the next newly captured one-line image data is input (step S6), and the process returns to step S3.

As described above, according to the first embodiment, the following effects can be obtained.
(1) Overhead inspection by prediction using image information of the previous time is possible.
(2) Overhead inspection using the overhead line information measured in the past as a state transition equation is possible.
(3) Overhead inspection using prediction based on probability theory is possible.

(2) Basic concept (Example 2)
An object of the present invention is to provide an overhead line inspection device that measures the displacement, overhead line width, and wear width of a plurality of overhead lines with high accuracy by image processing using a time series filter.

The difference from the first embodiment is that a plurality of overhead lines can be handled instead of a single overhead line.
In Example 1, since the overhead wire displacement, the overhead wire width, and the wear width are simultaneously obtained, each particle holds a three-dimensional parameter. However, when there are a plurality of overhead wires, it is necessary to change this parameter.

  Specifically, when there are two overhead wires 1a and 1b, the displacement of the overhead wire 1a, the width of the overhead wire 1a, the wear width of the overhead wire 1a, the distance between the overhead wire 1a and the overhead wire 1b (the displacement of the overhead wire 1a and the overhead wire 1b). Deviation), the width of the overhead wire 1b, and the wear width of the overhead wire 1b.

Using information on the widths of two overhead lines, which are known to some extent in advance, not only enables more accurate recognition than if there was only one overhead line, but also when the number of overhead lines changes It becomes possible to respond flexibly by changing the dimension of the particles.
Likelihood calculation is realized by associating the one used in Example 1 with the number of all overhead lines.

An apparatus configuration example according to the present embodiment is the same as that of the first embodiment as shown in FIG.
Therefore, as shown in FIG. 3, the overhead line inspection device according to the present embodiment includes an image input unit 10, a storage unit 20, a particle initialization unit 30, a likelihood calculation unit 40, a resampling unit 50, and a state transition unit. 60.

In the image input unit 10, image data for one line captured in time series by the line sensor is input and stored in the storage unit 20. The image data includes time-series overhead wire displacement, overhead wire width, and wear width as shown in FIG.
The storage unit 20 stores various parameter data such as image data, particle data, and state transition probabilities.

The particle initialization unit 30 sets random parameters (overhead deflection, distance of multiple overhead lines, overhead line width for the number of overhead lines, wear width for the number of overhead lines) in the range and density set for a certain number of particles. The range, density, and parameters of each particle are stored in the storage unit 20.
The likelihood calculation unit 40 calculates the likelihood using the parameter and image data of each particle based on the likelihood function, and stores the likelihood given to each particle in the storage unit 20.

The resampling unit 50 performs resampling so that particles having a high likelihood are selected with a high probability based on the likelihood.
The resampled particles are stored in the storage unit 20, and the particles used for the resampling are deleted.

  The state transition unit 60 changes the parameters of the resampled particles based on the state transition model, creates new particles, and stores them in the storage unit 20. According to the newly created particles, the deviation, overhead wire width, and wear width of a plurality of overhead wires are measured with high accuracy as the probability theory.

A flowchart according to this embodiment is shown in FIG.
First, one line of image data captured in time series by the line sensor is input by the image input unit 10 (step S1).
Next, particles having a random value within the range and density set for each parameter (deviation of a plurality of overhead lines, overhead line width, wear width) are prepared by the particle initialization unit 30 (step S2).

Subsequently, the likelihood calculation unit 40 calculates the likelihood of each particle using the parameter of each particle and the image data of the previous line based on the likelihood function (step S3).
Then, a resampling process is performed by the resampling unit 50 based on the likelihood, and a new particle is created by the state transition unit 60 by using the state transition model for the resampled particles (step S4).

  Thereafter, until the time-series imaging by the line sensor is completed (step S5), the next newly captured one-line image data is input (step S6), and the process returns to step S3.

As described above, according to the second embodiment, the following effects can be obtained.
(1) Multiple overhead inspection by prediction using image information of previous time is possible.
(2) Multi-line inspection using multi-line information measured in the past as a state transition equation is possible.
(3) Multiple overhead inspection using prediction based on probability theory is possible.

(3) Basic concept (Example 3)
An object of the present invention is to provide an overhead wire / suspended wire type inspection device that measures the displacement / wear width of a plurality of overhead wires and the displacement of a suspended wire with high accuracy by image processing using a time series filter.

The difference from the first and second embodiments is that it can be applied not only to the overhead line but also to the suspended overhead line. As in the first and second embodiments, the particle parameters are obtained by adding three parameters for one overhead line, two parameters for the suspension line displacement (relative distance from the overhead line), and the suspension line width for each suspension line. In addition, the suspension wire does not wear because it is not in contact with the pantograph. Therefore, there is no wear width on the suspension line.
Likelihood calculation is performed by adding a likelihood function to the likelihood used in Example 2 in consideration of the tendency that the pixel value of the suspension line tends to be lower than the background. That is, the likelihood function is also based on the difference between the pixel values of the background and the suspension line.

An apparatus configuration example according to the present embodiment is the same as that of the first embodiment as shown in FIG.
Therefore, as shown in FIG. 3, the overhead line inspection device according to the present embodiment includes an image input unit 10, a storage unit 20, a particle initialization unit 30, a likelihood calculation unit 40, a resampling unit 50, and a state transition unit. 60.

In the image input unit 10, image data for one line captured in time series by the line sensor is input and stored in the storage unit 20. The image data includes time-series overhead wire displacement, overhead wire width, and wear width as shown in FIG.
The storage unit 20 stores various parameter data such as image data, particle data, and state transition probabilities.

  In the particle initialization unit 30, random parameters (range of the overhead wire, distance of the plurality of overhead wires, overhead wire width for the number of overhead wires, wear width for the number of overhead wires, and the number of wires are set in a range / density set for a certain number of particles. The deviation of the suspension line and the width of the suspension line for the number of the suspension lines are set, and the range, density, and parameters of each particle are stored in the storage unit 20.

The likelihood calculation unit 40 calculates the likelihood using the parameter and image data of each particle based on the likelihood function, and stores the likelihood given to each particle in the storage unit 20.
The resampling unit 50 performs resampling so that particles having a high likelihood are selected with a high probability based on the likelihood.

The resampled particles are stored in the storage unit 20, and the particles used for the resampling are deleted.
The state transition unit 60 changes the parameters of the resampled particles based on the state transition model, creates new particles, and stores them in the storage unit 20. According to the created new particles, the deviation, overhead line width, wear width of the plurality of overhead wires and the width and deviation of the suspension wires are measured with high accuracy as the probability theory.

A flowchart according to this embodiment is shown in FIG.
First, one line of image data captured in time series by the line sensor is input by the image input unit 10 (step S1).
Next, set each parameter (overlapping wire displacement, distance of multiple overhead wires, overhead wire width for the number of overhead wires, wear width for the number of overhead wires, deflection of the suspension wire for the number of wires, suspension wire width for the number of wires) Particles having random values within the range and density are prepared by the particle initialization unit 30 (step S2).

Subsequently, the likelihood calculation unit 40 calculates the likelihood of each particle using the parameter of each particle and the image data of the previous line based on the likelihood function (step S3).
Then, a resampling process is performed by the resampling unit 50 based on the likelihood, and a new particle is created by the state transition unit 60 by using the state transition model for the resampled particles (step S4).

  Thereafter, until the time-series imaging by the line sensor is completed (step S5), the next newly captured one-line image data is input (step S6), and the process returns to step S3.

As described above, according to the third embodiment, the following effects can be obtained.
(1) Overhead / suspended line inspection by prediction using image information of previous time is possible.
(2) Overhead line / suspended line inspection using the information of the overhead line / suspended line measured in the past as a state transition equation is possible.
(3) Overhead / suspended line inspection using prediction based on probability theory is possible.

(4) Basic concept (Example 4)
An object of the present invention is to provide an overhead wire / suspension wire / crossover wire inspection device that accurately measures the displacement / wear width / suspension wire / shift wire displacement of a plurality of overhead wires by image processing using a time series filter. It is to be.

The difference from the third embodiment is that it can cope with not only an overhead line and a suspended overhead line but also a crossover line.
The particle parameters are the same as in Example 3. Three parameters for each overhead wire, two parameters for the suspension wire displacement (relative distance from the overhead wire) and the suspension wire width for each suspension wire, crossover wire displacement, and jumper wire width These two parameters are added. Note that the suspension and connecting wires do not wear because they are not in contact with the pantograph. Therefore, there is no wear width in the suspension wire and the crossover wire.
Likelihood calculation is performed by adding a likelihood function that takes into account that the pixel value of the crossover line tends to be lower than the background to the one used in the third embodiment. That is, the likelihood function is also based on the difference between the pixel values of the background and the crossover.

An apparatus configuration example according to the present embodiment is the same as that of the first embodiment as shown in FIG.
Therefore, as shown in FIG. 3, the overhead line inspection device according to the present embodiment includes an image input unit 10, a storage unit 20, a particle initialization unit 30, a likelihood calculation unit 40, a resampling unit 50, and a state transition unit. 60.

In the image input unit 10, image data for one line captured in time series by the line sensor is input and stored in the storage unit 20. The image data includes time-series overhead wire displacement, overhead wire width, and wear width as shown in FIG.
The storage unit 20 stores various parameter data such as image data, particle data, and state transition probabilities.

In the particle initialization unit 30, random parameters (range of the overhead wire, distance of the plurality of overhead wires, overhead wire width for the number of overhead wires, wear width for the number of overhead wires, and the number of wires are set in a range / density set for a certain number of particles. The displacement of the suspension line, the width of the suspension line, the displacement of the connection line, and the width of the connection line) are set, and the parameters of the range, density, and each particle are stored in the storage unit 20.
The likelihood calculation unit 40 calculates the likelihood using the parameter and image data of each particle based on the likelihood function, and stores the likelihood given to each particle in the storage unit 20.

The resampling unit 50 performs resampling so that particles having a high likelihood are selected with a high probability based on the likelihood.
The resampled particles are stored in the storage unit 20, and the particles used for the resampling are deleted.
The state transition unit 60 changes the parameters of the resampled particles based on the state transition model, creates new particles, and stores them in the storage unit 20. According to the newly created particles, the deviation, overhead wire width, wear width of multiple overhead wires, suspension wire width, deflection overhead wire displacement, and crossover deflection / width were measured with high accuracy as the probability theory. become.

A flowchart according to this embodiment is shown in FIG.
First, one line of image data captured in time series by the line sensor is input by the image input unit 10 (step S1).
Next, parameters (overlapping wire displacement, distance of multiple overhead wires, overhead wire width for the number of overhead wires, wear width for the number of overhead wires, deflection of the suspension wire for the number of wires, suspension wire width for the number of wires, crossover wire Particles having a random value in the range and density set in (deviation of the line, width of the connecting line) are prepared by the particle initialization unit 30 (step S2).

Subsequently, the likelihood calculation unit 40 calculates the likelihood of each particle using the parameter of each particle and the image data of the previous line based on the likelihood function (step S3).
Then, a resampling process is performed by the resampling unit 50 based on the likelihood, and a new particle is created by the state transition unit 60 by using the state transition model for the resampled particles (step S4).

  Thereafter, until the time-series imaging by the line sensor is completed (step S5), the next newly captured one-line image data is input (step S6), and the process returns to step S3.

As described above, according to the fourth embodiment, the following effects can be obtained.
(1) Overhead / suspended / crossover lines can be inspected by prediction using image information of the previous time.
(2) Overhead line, suspension line, and crossover line measurement using the information of overhead lines, suspension lines, and crossover lines measured in the past as state transition equations is possible.
(3) Overhead, suspended and crossover inspections using prediction based on probability theory are possible.

(5) Basic concept (Example 5)
It is an object of the present invention to accurately detect the deviation, wear width, suspension line, crossover line, and height of each overhead line by image processing using a plurality of (two or more) line sensor cameras and a time series filter. It is to provide an overhead line / suspended line / crossover line inspection device for measuring.

The difference from Example 4 is that it is possible to cope with the height of overhead lines, suspension lines, and crossovers by preparing multiple cameras (the height is determined by triangulation method). This is the point that can be achieved.
The parameter of the particle is the number of cameras of the parameter of the fourth embodiment.
Likelihoods used in the fourth embodiment are used for the number of cameras.

An apparatus configuration example according to the present embodiment is the same as that of the first embodiment as shown in FIG.
Therefore, as shown in FIG. 3, the overhead line inspection device according to the present embodiment includes an image input unit 10, a storage unit 20, a particle initialization unit 30, a likelihood calculation unit 40, a resampling unit 50, and a state transition unit. 60.

In the image input unit 10, image data for one line captured in time series by the line sensor is input and stored in the storage unit 20. The image data includes time-series overhead wire displacement, overhead wire width, and wear width as shown in FIG.
The storage unit 20 stores various parameter data such as image data, particle data, and state transition probabilities.

  In the particle initialization unit 30, the parameters for the number of random cameras in the range and density set for a certain number of particles (overhead wire displacement, distance of multiple overhead wires, overhead wire width for the number of overhead wires, wear width for the number of overhead wires) , The number of suspension lines as many as the number of suspension lines, the width of the suspension lines as many as the number, the displacement of the connection lines, and the width of the connection lines) are set, and the parameters of the range, density, and each particle are stored in the storage unit 20.

The likelihood calculation unit 40 calculates the likelihood using the parameter and image data of each particle based on the likelihood function, and stores the likelihood given to each particle in the storage unit 20.
The resampling unit 50 performs resampling so that particles having a high likelihood are selected with a high probability based on the likelihood.
The resampled particles are stored in the storage unit 20, and the particles used for the resampling are deleted.

  The state transition unit 60 changes the parameters of the resampled particles based on the state transition model, creates new particles, and stores them in the storage unit 20. According to the created new particles, as a probability theory, the displacement, wear width, suspension line, crossover line, and height of each of the plurality of overhead lines are measured with high accuracy.

A flowchart according to this embodiment is shown in FIG.
First, one line of image data captured by a plurality of line sensor cameras in time series × the number of cameras is input by the image input unit 10 (step T1).
Next, particles having a random value in a range and density set for each parameter (deviation, wear width, suspension line, crossover line, and height of a plurality of overhead lines) corresponding to the number of cameras are set to a particle initialization unit 30. (Step T2).

Subsequently, the likelihood calculation unit 40 calculates the likelihood of each particle using the parameter of each particle and the image data of the previous line based on the likelihood function (step T3).
Then, a resampling process is performed by the resampling unit 50 based on the likelihood, and a new particle is created by the state transition unit 60 by using the state transition model for the resampled particles (step T4).

  Thereafter, until the time-series imaging by a plurality of line sensors is completed (step T5), the next newly captured image data of one line × the number of cameras is input (step T6), and the process returns to step T3. .

As described above, according to the fifth embodiment, the following effects can be obtained.
(1) Overhead / suspended / crossover lines can be inspected by prediction using image information of the previous time.
(2) Overhead line, suspension line, and crossover line measurement using the information of overhead lines, suspension lines, and crossover lines measured in the past as state transition equations is possible.
(3) Overhead / suspended / crossover lines can be detected using prediction based on probability theory.
(4) It is possible to increase the accuracy of overhead / suspended / crossover line inspection using multiple cameras and to measure the height of overhead / suspended / crossover lines.

  INDUSTRIAL APPLICABILITY The present invention can be widely used industrially as an overhead wire inspection device for inspecting and measuring overhead wire displacement using a time series filter.

DESCRIPTION OF SYMBOLS 10 Image input part 20 Memory | storage part 30 Particle initialization part 40 Likelihood calculation part 50 Resampling part 60 State transition part

Claims (6)

In an overhead line inspection device that accurately measures overhead lines on a train by image data processing using a probabilistic time series filter,
An image input unit for inputting image data including a deviation, a width, and a wear width of the overhead wires taken in time series by a line sensor disposed on the upper surface of the train;
A particle initialization unit for preparing particles having random parameters of deviation, width and wear width of the overhead wires,
A likelihood calculating unit that calculates a likelihood value based on a likelihood function using the image data of the previous time input by the particle prepared by the particle initialization unit and the image input unit;
Based on the likelihood value calculated by the likelihood calculation unit, a resampling unit that performs resampling so that particles with a high likelihood value are selected with a high probability;
An overhead line inspection and measurement apparatus comprising: a state transition unit configured to change a parameter of the particles resampled by the resampler based on a state transition model to create a new particle.   The overhead line inspection device according to claim 1, wherein the overhead lines are single or a plurality of overhead lines.   The overhead line inspection device according to claim 2, wherein the overhead lines include a suspended overhead line.   The overhead line inspection device according to claim 2 or 3, wherein the overhead lines include crossovers.   2. The overhead lines according to claim 1, wherein the number of the line sensors is plural, the height of the overhead lines is also measured, and the particle initialization unit includes the height of the overhead lines as a parameter. Inspection equipment.   The likelihood function is an absolute value of a difference between pixel values before and after the wear area of the overhead line, and a difference value when the average pixel value of the overhead line area is subtracted from an average pixel value of the wear area of the overhead line. The overhead wire inspection device according to claim 1, wherein the overhead wire inspection device is based on a deviation, width, and wear width of the overhead wire at a previous time and a difference between the displacement, width, and wear width of the overhead wire at a current time. .