Hough transform and thermo-vision for monitoring ...webm.dsea.unipi.it/~saniw/public_html/Bari/P_10.pdf · Obtained results are relevant for monitoring the status of the quality transmission

Hough transform and thermo-vision for monitoringpantograph–catenary systemA Landi�, L Menconi, and L Sani

Dipartimento di Sistemi Elettrici e Automazione, Universita di Pisa, Pisa, Italy

The manuscript was received on 2 November 2005 and was accepted after revision for publication on 17 July 2006.

DOI: 10.1243/0954409JRRT41

Abstract: The pantograph-overhead contact wire system is investigated by using an infraredcamera. As the pantograph has a vertical motion because of the non-uniform elasticity of thecatenary, in order to detect the temperature along the strip from a sequence of infraredimages, a segment-tracking algorithm, based on the Hough transformation, has been employed.An analysis of the stored images could help maintenance operations revealing, for example,overheating of the pantograph strip, bursts of arcing, or an irregular positioning of the contactline. Obtained results are relevant for monitoring the status of the quality transmission of thecurrent and for a predictive maintenance of the pantograph and of the catenary system.Examples of analysis from experimental data are reported in the paper.

Keywords: catenary/pantograph system, Hough transform, high-speed trains, thermo-vision

1 INTRODUCTION

The development of innovative methods for improv-ing efficiency, reliability, and safety in railway systemis a primary objective of the railway research.Especially in the case of high-speed trains, a regularcurrent transmission in pantograph–catenary sys-tems becomes a very difficult task to be satisfied. Aregular current transmission in railway systems ismore and more important if the European Standard96/48/CE about interoperability is considered: Stan-dard 96/48/CE defines the rules for an efficient trans-national circulation of high-speed trains in Europe,i.e. a locomotive will be designed to freely operateinside European countries independently from thefeeding voltage and frequency. It is well known thata high quality current collection is characterized bya continuous contact between the pantograph andthe overhead contact wire. A poor contact producesvarious drawbacks, including break arcs. If breakarcs have a long duration, the efficiency of the loco-motive may be reduced, and moreover an excessivewear of the pantograph strips and of the contact

wire may be produced. Besides the acceleratedwear of materials and the electrical problems, themost harmful and dangerous effect due to a defectivecurrent collection is an unexpected sudden breakingof the contact wire: such event can put out of servicelines and trains for hours or days and can cause acci-dents with possible threat to passengers’ health.Total heat quantity dissipated by the sliding contactduring the current transmission is the sum ofthe heat by friction, contact resistance, and arcdischarge, generated when the contact breaks.Temperature of arc plasma is known to reach3500–4000 K, therefore, temperature of the surfaceregion rises promptly around the spot where thearc is discharged. In case of bursts of arcing, elec-tronic regulators (usually switching regulators)suffer anomalous stresses, consequently high-orderharmonics on the line current further deterioratethe efficiency and reduce the quality of the tractiontorque. Furthermore, the presence of harmonicsmay be dangerous for possible interactions with sig-nalling transmissions, with a critical increasing ofelectro magnetic emission and a dangerousreduction of safety. The risk of harmful interferencephenomena is particularly relevant when also thesignalling systems employ the tracks as a part oftheir circuitry, which was in the past a useful techni-cal solution to reduce systems build-up costs.

�Corresonding author: Dipartimento di Sistemi Elettrici e

Automazione, University of Pisa, Via Diotisalvi 2, Pisa 56126,

Italy. email: [email protected]; [email protected]

435

JRRT41 # IMechE 2006 Proc. IMechE Vol. 220 Part F: J. Rail and Rapid Transit

Therefore, a reliable detection of the break arcs isvery important for setting up controlled systems(e.g. active pantographs), for reducing the acceleratewear of the materials and for avoiding as much aspossible any interference with communication net-works. High speeds worsen these negative aspects:the higher speeds, the more critical the problems.Over 200 km/h, wires start oscillating and the panto-graph head cannot maintain contact in the case ofabrupt height variations. The presence of morethan one pantograph on a train worsens the problemof oscillating wires for the rear pantograph. Itencounters an overhead line already excited by thefirst one, so that the quality of the contact deterio-rates and break arcs increases their frequency andduration. In order to improve the current collectionat high speed and to reduce the deterioration dueto break arcs, different solutions have been pro-posed. The traditional solution consists of increasingthe uplift contact force between the pantograph andthe contact wire, at the expense of a reduction of thelife-time of the collector strips and particularly of thecontact wire, because of erosive and abrasive wearand of temperature raising. Such a solution wasadopted in the past, but it is today inadequate forhigh speed trains. A primary objective in railwayresearch is the reduction of contact forces. A firstinteresting solution is the design of innovative panto-graphs with reduced head mass and active control.All railway companies are well aware of the problemsaffecting the current collection: a relevant objectiveboth for academic researchers and for railway com-panies is the development of innovative pantographsand the application of new sensors for a continuousmonitoring of the sliding contact and for helpingmaintenance activities. Therefore, many universitieshave devoted their activities to deal problems relativeto the rail transportation.

The main result of our past research [1, 2] was thestudy and implementation of sensors (UV and cur-rent sensors) for detecting break arcs. A correlationof measured repetitive break arcs with the kilometricprogression of the railway line Rome–Florencerevealed its validity for an automatic checking ofthe status of the contact overhead line. These resultsvalidated from a strict cooperation with Trenitalia(Italian Railways) proved to be a useful tool for help-ing maintenance activities both for the overhead lineand for the pantograph.

The new research line proposed is based on thestudy of the quality of the contact pantograph–catenary with infrared techniques. The researchbased on thermography is totally new for pantographstudies and its results are foreseen to be meaningfulfor testing new materials for collector strips and newtypologies of pantographs and catenary. Thermo-graphic images are more informative than standard

camera images, because of their insensitivity todifferent weather conditions, to the daily-nightlyruns or to the presence of tunnels. Hot spots on theoverhead line and on the pantograph strips are easyto detect and a careful analysis of the stored imagescould help maintenance operations in the case ofirregular positioning of the line (e.g. the wireshould be staggered to even the wear on the train’spantograph: in the case of a defective layout of theline, the pantograph strip dramatically overheatsand such effect should be easily revealed).

2 THERMOGRAPHY FOR MONITORINGPANTOGRAPH–CATENARY INTERACTIONS

Thermo-cameras have not been used extensively inrailway research. In reference [3] preliminary resultsbased on the analysis of infrared camera acquisitionswere proposed. The quality of the current collectionwas evaluated by monitoring the infrared emissionat the contact point. Measurements carried outduring high-speed test runs and a post processing ofthe data collected have revealed the effectiveness ofthe proposed method for the detection of the lossesof contact, in order to evaluate and test the perform-ance of running pantographs and of the contact wires.

Main advantages offered by thermo-vision are asfollows.

1. It is a non-contact and non-destructive technique.2. It is suited to monitor devices operating under

high voltage or carrying high currents.

A critical aspect of thermo-vision is that the exacttemperature of the body under test cannot bedirectly revealed. Nevertheless infrared imaging isan excellent method for extracting a qualitativemap of the superficial temperature. As a matter offact, thermal analysis gives relative information onthe temperature, quoting [4]: ‘all objects at tempera-tures above absolute zero emit electromagnetic radi-ation. Radiation thermometry makes use of this factto estimate the temperatures of objects bymeasuringthe radiated energy from selected regions’.

Each physical process characterized by an increaseor decrease in surface temperature is detectable withinfrared thermography. The intensity of the emittedradiation depends on two factors: the body tempera-ture and the factor emissivity of the surface, i.e. theability of the object to radiate, defined by theStefan–Boltzmann equation

E ¼ 1sT 4 (1)

The emissivity factor 1 lies between zero and theunity, depending on the material nature and on thesuperficial roughness. Infrared imaging systems

436 A Landi, L Menconi, and L Sani

Proc. IMechE Vol. 220 Part F: J. Rail and Rapid Transit JRRT41 # IMechE 2006

convert infrared heat emissions into a picture show-ing the relative temperature differences in a range ofgrey tones, or in a series of colours, in such a way thatthe desired temperature information is easily inter-preted by the user. An analysis of the acquiredimages leads to very interesting results: an IR ima-ging is more informative than a standard cameraimage.

In Fig. 1(a), a typical infrared image of panto-graph–catenary interaction is shown (the trainspeed was 200 km/h in this case). The maximuminfrared emission is in the neighbourhood of thecontact surface between strip and wire. The inter-action region between the overhead feeder and thestrip shows a high thermal gradient. A continuousmonitoring of the contact region gives interestinginformation on the quality of the current trans-mission: a continuously increasing temperaturewith slow time constant denotes that friction isincreasing. On the other hand, fast phenomena pro-duce bursts of arcing, i.e. losses of contact betweencatenary and pantograph (Fig. 1(b)).

A first question for applying thermo-cameraimages to pantograph–catenary interaction is: howto follow the contact point? In fact the wire is stag-gered relatively to the centre-line of the track, toeven the wear on the train’s pantograph as it runsunderneath. Moreover, the height of the contactwire may change (e.g. entering a tunnel). Therefore,a relevant displacement of the contact point maycomplicate the image processing, unless themoving spatial coordinates of the contact area areconsidered as known. A first step for an automaticimage processing analysis is to follow frame-by-frame the position of some critical features (panto-graph strips, overhead electric line). All theseelements can be characterized by linear contours.This observation allows using standard algorithms

for detecting straight lines inside each image underexamination.

3 APPLICATION OF THE HOUGH TRANSFORMFOR MONITORING PANTOGRAPH–CATENARYINTERACTION

The task of the proposed research is the applicationof tracking algorithm of predefined objects (e.g. thecontact point on the pantograph strip) over asequence of images acquired from thermo-camera,in order to monitoring variables of interest (e.g.temperature at the contact point). The Hough trans-form is well suited for detecting straight-lines in thereal image: its application can be extremely usefulfor monitoring the contact point, characterizedfrom the intersection of a segment representing thestrip and a straight line representing the overheadline contact. Furthermore, it can identify differentrectilinear structures, such as the poles along therailway line, helping in finding a more precise corre-lation between the overheating of the contact point(or a burst of arcing) and the position of the train.A detailed description of the Hough [5] transform isreported in Appendix 2.

The analysis presented in this paper considersimages recorded along the Florence–Rome railwayline. Suppose to know approximately the positionand the possible movement of the features to bemonitored, so that two observation windows intothe image are locked: a first one includes the panto-graph structure and the second one reveals the pas-sage of the support structures.

Independently from the electronic format of therecorded file (outputs are considered in .bmp, .mpgor .avi formats), single frames are extracted fromthe image sequence. Each image is first stored in a

Fig. 1 (a) Infrared image of pantograph–catenary interaction and (b) thermal image in case of a

burst of arcing

Monitoring of pantograph –catenary system 437


matrix, then converted from true colour (RGB) togrey scale.

Each image is suitably filtered for reducing noisesand sharpens objects, then edges are extracted andimage is then converted from a grey scale to abinary image. Edge detection was performed usingthe Canny algorithm [6] with a preliminary noisereduction.

After this step, each frame is represented by abinary image with edges evidenced, making theidentification of the pantograph, the overhead con-tact lines, and the poles, an easy task.

Object locations in each frame is characterizedfrom an analysis of edges and a detection of refer-ence straight lines.

All straight lines can be localized using the Houghtransform: to increase algorithm efficiency and toreduce computational burden, the straight linesearch can be limited inside suitable rectangularregions of the image. Sizes of the windows andtheir positions relative to the full image are choseninside the sequence: the hypothesis is that theyhave constant dimensions inside the sequence.

A relevant advantage of reducing the analysis to awindow inside the full image, is that computationaltime may be greatly reduced and a good practicalchoice is to impose simple rules to define the pos-itions of the windows as soon as possible.

An object recognizing procedure allows a fastdetection of the object position and a repetitivealgorithm allows following the trajectory of anobject inside successive frames. After a first detectionof the trajectory, it is a good strategy to eliminateframes with discontinuous trends of the object andto filter the image for reducing noise and to sharpenthe image.

Following this logic, a software tool was createdthat allows the user to automatically detect:

(a) pantograph position;(b) catenary wires;(c) position of the poles.

This procedure can be repeated for each frameof the sequence. As the pantograph moves slowlywith respect to the frame rate of the camera,the position of the strip detected is used in thelast image to window next image appropriately.Therefore, the computation time can be greatlyreduced.

3.1 Position of the pantograph inside an image:detection logic

The pantograph is the key object of the recordedscene: it appears at each frame and its positionmoves slowly. Small displacements both in vertical

and in horizontal direction are allowed: they aredue to the movements of the main frame and ofthe contact arms.

In thermo-images, two different pantograph headsand two different strips are evidenced: the goal of thefollowing study is to follow only the strip nearest theIR camera. To be clearer, the task is to detect andfollow the nearest pantograph head, for monitoringvariations in the temperature of the upper edge, i.e.in the contact strip.

By considering the application of the Cannyalgorithm for edge detection (Figs 2(a) and (b)),some almost horizontal segments are detected (evi-denced inside a red ellipse in Figs 2(c) and (d) inthe following study they will be considered as refer-ence lines).

Contact analysis is applied to the squared windowof Fig. 3(a), named reference box: its size and pos-ition inside the image will remain unchanged forthe limited displacement of the pantograph head.Several segments are easily recognized by Houghtransform (Fig. 3(a)) only the upper one representsthe contact line for monitoring (Fig. 3(b)).

Reference box allows following the pantographdisplacements along a vertical direction. For a cor-rect analysis, the horizontal displacement also hasto be measured. Lateral displacements can then beestimated from an analysis of the two lateral armssustaining the pantograph heads, (Fig. 3(c)).

The two segments representing the external profileof these lateral supports can be identified by repeat-ing the Hough analysis inside two new referenceboxes, as shown in Fig. 3(d)).

Fig. 2 Edge detection: (a) Canny algorithm, (b) Canny

algorithm with noise reduction, (c) pantograph

edge detection, and (d) support infrastructure

and overhead wires edge detection



The three reference boxes can be joined togetherfor localizing a triangle as the reference shape repre-senting the pantograph, as shown in Fig. 4.

The upper side of the triangle is representingthe line useful for monitoring the contact with thecatenary wires. Note that the representation of thepantograph head with a straight line is only anapproximation of the real scenario. Nevertheless,the maximum error accumulates in the lateralregions of the segment, where the pantograph headis not rectilinear: as a matter of fact in such regions

an analysis of the contact is less relevant, as thelocal temperature is lower than in the middle zoneof the segment.

3.2 Identification of the contact betweenpantograph and overhead contact wire

After the identification of the pantograph structure,the following step is devoted to the identificationof the contact zone between pantograph and the

Fig. 3 Application of the Hough transformation: (a) horizontal detected segments into the

reference box of the strip, (b) upper edge of the nearest strip, (c) detected segments into

the reference boxes of the lateral supports, and (d) detected lateral supports

Fig. 4 Pantograph detection: (a) detected line by Hough transformation and (b) reference shape



overhead wire. In the case of the image sequencesrecorded along the railways from Florence to Rome,it can be observed that there are many straightlines, between them two overhead wires (Fig. 5(a))are in close proximity to the contact region.

It must be highlighted that the contact points arenot corresponding to the hottest points of the strip,usually in themiddle of the contact strip. Such obser-vation excludes to follow maxima temperatures fordetecting the contact points and it requires consider-ing the interaction between the upper side of the tri-angle representing the pantograph and the overheadcontact wires.

Hough transform can be applied again to detectoverhead wires, as shown in Fig. 5(b).

As shown in Fig. 5, many straight lines aredetected. A selection of the correct contact wireswas performed on the basis of the two following rules.

1. Contact wires are always adjacent and at theexterior of the set of the identified lines.

2. Contact wires are thicker than non-contact ones.

After the detection of the pencil of lines of Fig. 5(b),the contact points are uniquely located via thefollowing steps.

1. Computing the intersectionbetween the upper sideof the triangle (pantograph head) and the twostraight lines exterior to thepencil of lines (Fig. 5(c)).

2. Analysing the colour distribution of the pixelsinside the image, near the two points selected atstep 1. The thickest line of the two is selected,for detecting the contact zone, whose extensionis predefined by a fixed number of pixels in thedirection of the middle of the pantograph head(Fig. 5(d)).

In such a way, only a reduced box surrounding thecontact points is recognized, instead of the two con-tact points; nevertheless, this result constitutes agood approximation both for monitoring the temp-erature in the contact region (e.g. its mean valueinside the reduced box) and for checking a correctpositioning of the contact wire in terms of its lateralmovements, relatively to the centre-line of the track.

Fig. 5 Identification of the contact between pantograph and overhead contact wire: (a) image

under examination, (b) detected straight lines representing catenary wires, (c) detected

boundaries, and (d) identified contact region



3.3 Passages under the portal structuresof the railways

In the case of a railway line with portal structures, asimilar method can be employed for detecting thepoles. The horizontal structure is the easiest elementto be recognized: it is the wider structure and itappears even if the poles are not included insidethe image, as shown in Fig. 6(a).

Therefore, by counting the passages of horizontalportals, it is easy to localize the train position alongthe line: this task can be performed using an obser-vation box positioned in a suitable region of theimage. An easy solution is to locate this box in aregion where only the horizontal portals are passing,e.g. in the upper left side of each frame (Fig. 6(b)).

The passages of horizontal portals must bedetected avoiding any possible false condition. TheHough transform for straight lines recognizing canbe applied inside the observation box and a reliablecriterion of portal detection may be based on anaccurate analysis of the accumulation matrix. Itsmaximum value is, very high in case of passages ofportal structures: a threshold value is therefore,useful for discriminating noisy images or for singlestraight lines passages.

In Fig. 7, the values contained in the accumulatormatrix for a sequence of 500 frames are shown: thedifferences of maxima in cases of passages of portalstructure are clearly evidenced.

Cases of different consecutive values overcomingthe chosen threshold in adjacent frames are due tothe finite time of the portals for crossing the obser-vation box: a correct countingmust take into accountonly the first of a set of consecutive frames. A variabletime delay, depending on the speed of the train, canbe introduced for evaluating the passages under-neath the portal.

To be clearer, consider an IR camera-fixed refer-ence frame: the relative motion of the portal has an

opposite direction with respect two the train, asdepicted in Fig. 8(a). The grey triangle is representingthe vision angle of the IR camera, the black square isthe location of the portal at the instant when the pan-tograph is passing underneath. In such instant, theportal is outside the vision angle and only when thetrain moves, the portal is revealed from the cameraimage (Fig. 8(b)). The distance s between the pointsA and B is constant and can be computed from theknowledge of the vision angle and on its slope withrespect to the horizontal axis. If the train speed isacquired, it is an easy computation to obtain thetime delay of the portal for moving from point A topoint B, and it is possible to estimate the exact timeof the passage of the pantograph under the portalstructure from the knowledge of the instant whenthe portal is revealed from the camera.

3.4 Trajectory improvement

All the procedures previously described lead to areliable recognizing of the pantograph and of thecontact region included into the IR images: they arerepeated for each frame of the overall sequence.

Fig. 6 (a) Horizontal bridge sustaining catenary wires, and (b) observation box for counting the

passages of portals

Fig. 7 Maxima contained in the accumulator matrix

for a sequence of 500 frames



The procedures work correctly, apart from a veryrestricted number of cases, where the algorithmfails or the results are not accurate. In Fig. 9, the tra-jectory of the contact point is shown for a sequenceof 20 frames.

Corresponding to frames 11 and 13, two outlierpoints are detected: each abrupt variation of the tra-jectory is due to errors in the recognizing procedurebecause of the physical continuity of the contactpoint in adjacent frames. This problem is easilysolved using a threshold for discriminating unaccep-table variations of the position in adjacent framesand, in case of outlier points, replacing them withdifferent values (e.g. the previous value can besaved and maintained).

To smoothen the scattering effect of the obtainedpoints, a low-pass filtering action can be introduced,e.g. using a first-order filter of Bessel type (Fig. 9(b)).The cut-off frequency of the filtering action can besuitably chosen as a function of the velocity of displa-cement of the objects in the sequence.

Circles are the original values, straight lines rep-resent the sequence after the outlier replacement,and dotted line is the result of the first-order filtering

action: it is a realistic smooth approximation of thecontinuous displacement of the contact point.

4 EXPERIMENTAL RESULTS

An accurate choice of the infrared camera is a veryimportant factor. Two categories of camera (IR) arecommercially available: short wavelengths infraredcameras (SW_IR) 3–5 mm and long wavelengthsinfrared cameras (LW_IR) 8–13 mm. Infraredimages have to look through atmosphere, but,unfortunately, the atmosphere does not transmitinfrared energy well at all wavelengths. SW andLW are the spectral ranges with high atmospherictransmission factor. Short wave systems are not asgood for outdoor use, because of greater interfer-ences from natural sunlight. Long wave systemsare also less distorted by the atmosphere: this is abasic motivation for employing them in ourresearch. A preliminary laboratory test was carriedout in order to yield the emissivity of the panto-graph strip. In Italian railways (3 kV d.c.), the con-tact material strip is copper (or copper based

Fig. 8 Portal structure displacement in an IR camera-fixed reference frame

Fig. 9 Trajectory of the contact point: (a) original (b) with replacement of outliers and filtering

action (dotted line)



alloys). By using the IR camera software, the aver-age emissivity of the strip was determined to be0.2. The measurement equipment was theninstalled on board of an ETR500, a high-speedtrain of Italian railways. The block diagram ofFig. 10 illustrates the measurement chain.

Trial runs have been carried out travelling alongthe railway line connecting Florence to Rome andvice versa. IR camera (produced by the Flir Systems)was located on the top deck of the locomotive, infront of the pantograph.

4.1 Portal structure detection of a positioningof the overhead contact line

In Figs 11(a) and (b), two plots are shown, the firstone consider a portal structure detection, and thesecond one reports the position of the contactpoint with respect to the centre of the pantographhead (a positive value means the contact point islocated on the right w.r.t. the median position).They are referred to an analysis of 200 frames. Plotsare functions of a position relative to the first frameanalysed. Train speed is considered as constantthat is equal to 200 km/h.

It can be noted that portal structures are posi-tioned each 60 m about, in a good agreement withthe expected layout. Each portal corresponds to aninversion of the direction of displacement of the con-tact point: it indicates a correct positioning of theoverhead line.

4.2 Example of break arc detection

Consider now, 100 consecutive frames with a breakarc corresponding to frame 50, shown in Fig. 12(a).The observation box is representative of the contactpoint.

The temperature profile along the strip surface forframe 50 is shown in Fig. 12(b). An analysis of theoverall frames gives the trend of temperatures; inFig. 13(a), the maximum temperature on the strip

Fig. 11 (a) Positioning of structural portal and (b)

contact point position respect to the centre

of the pantograph head

Fig. 12 (a) Frame 50 with break arc detection and (b)

temperature profile on the strip surfaceFig. 10 Measurement chain



surface is considered, in Fig. 13(b) the maximumtemperature in a region surrounding the contactpoint. Both of them indicate a burst of arcing.Figure 14 shows the infrared map of the strip alongthe railway line connecting Florence to Romebetween 60 and 63 km. Temperature peaks are corre-lated with arcing occurrences. This elaboration hasbeen obtained by processing 2700 frame of thevideo sequence with a computation time of 540 s(using an AMDw Athlon processor, 1.2 GHz).

5 CONCLUSIONS AND FUTURE WORK

The use of an infrared camera for monitoring thepantograph–catenary status has been proposedand experimentally tested. The temperature profileof the pantograph strip was detected on infraredimages by using the Hough transformation. Ananalysis of the stored images could help main-tenance operations, revealing, for example anirregular positioning of the line. In overhead contact

Fig. 13 (a) Maximum temperature on the surface of the strip and (b) maximum temperature in a

neighbourhood of the contact point

Fig. 14 Infrared map of pantograph strip along the railways



lines, the wires should be staggered, relative to thecentre-line of the track to even the wear: in thecase of a defective layout of the line, the pantographstrip dramatically overheats and a thermal analysiseasily reveals such effect.

Future work will consist of a comparative testing ofnew materials for collector strips. New typologies ofpantographs and catenary layout can be thermicallyanalysed.

ACKNOWLEDGEMENT

The financial support of the Italian Ministry ofUniversity and Research–MIUR (Project onImprovement of Current Transmission for HighSpeed Trains by means of Active Pantographs andNon-Invasive Diagnostics) is gratefully acknowl-edged. The authors wish to thanks P. Masini andM. Papi of the Italian railways for their kindcooperation.

REFERENCES

1 Bruno, O., Landi, A., Papi, M., and Sani, L. Phototubesensor for monitoring the quality of current collectionon overhead electrified railways. Proc. Instn Mech.Engrs, Part F: J. Rail and Rapid Transit, 2001, 215(F3),231–241.

2 Barmada, S., Landi, A., Papi, M., and Sani, L. Waveletmulti-resolution analysis for monitoring the occurrenceof arcing on overhead electrified railways. Proc. InstnMech. Engrs, Part F: J. Rail and Rapid Transit, 2003,217(F3), 177–187.

3 Balestrino, A., Bruno, O., Landi, A., Masini, P., Min-gozzi, E., Papi, M., and Sani, L. Infrared camera formonitoring pantograph–catenary interactions. Proceed-ings of World Conference of Railway Research 2003,Edinburgh, UK, 2003, pp. 1–7.

4 Runciman, H. M. Thermal imaging. In The measure-ment, instrumentation, and sensors handbook, (Ed. J. G.Webster), 1998 (CRC Press Inc., Boca Raton, FL).

5 Hough, P. Method and means for recognizing complexpatterns. US Pat. 3069654, 1962.

6 Deans, S. R. Hough transform from the Radon trans-form. IEEE Trans. Pattern Anal. Mach. Intell., 1981,PAMI-3(2), 185–188.

7 Russ, J. C. Image measurements. The image processingtoolbox, 2000 (CRC Press Inc., Boca Raton, FL).

8 Canny, J. A computational approach to edge detection.IEEE Trans. Pattern Anal. Mach. Intell., 1986, PAMI-8(6), 679–698.

APPENDIX 1

Notation

E radiation in W/m2

T temperature in Kelvin

1 emissivitys Boltzmann constant

APPENDIX 2

The Hough transform

In order to detect automatically the temperature ofthe contact strip, a tracking algorithm must beemployed. The problem is typical of the image analy-sis: it consists of detecting and tracking straight linesin a sequence of images. Several methods applied totwo-dimensional binary images are well-known inthe literature [4]: among them a powerful methodfor detecting a segment in a binary image is basedon the Hough transform [5]. Therefore, the problemaddressed can be summarized in the application of asuitable search algorithm for extracting edges fromthe binary image and for grouping pixels fromedges, such to identify (within a tolerance bound)collinear pixels.

In the sequel, a short description of the Houghtransform is presented.

Hough transform (1962) is a modification of theRadon transform, specialized to identify collinearpoints (lines) in digital images [6]. It was naturallyderived for identifying straight lines in a two-dimensional space for parameters, although itcan be extended to fit other regular shapes (circles,ellipses, squares, etc.,) at the expense of anincreasing dimensionality in the space of par-ameters, e.g. a circular Hough transform requires athree-dimensional space for parameters. A general-ized Hough transform can be employed in thecases where a simple analytical description ofshapes is not possible. The main feature of theHough transform is that it is an algorithmic pro-cedure based on a pixel image transformation intoa new space, denoted as Hough or parameters’space. After defining the line that needed to besearched into the image, every pixel of the edge indi-cates its contribution to the defined line: every pointpresent in the real space image casts its votes into theHough space for each of the lines that can possiblypass through it. To be clearer, consider the problemof straight line detection. The straight line connect-ing a sequence of pixels can be expressed in Carte-sian coordinates as

y ¼ mx þ b (2)

where the slope m and the intercept b are the twoparameters describing any straight line.



Unfortunately, Cartesian description has the advan-tage that in an image m can become infinitely largefor vertical lines. Therefore, the polar coordinatesare usually considered

r ¼ x cos (u)þ y sin (u) (3)

where (r, u) defines the shortest vector from origin tothe line, as shown in Fig. 15(a). This vector is perpen-dicular to the line. Consider now a two-dimensionalspace defined by two parameters r and u (Houghspace). Any line in the x–y plane plots to a point inthis space (Fig. 15(b)).

If a point (xi, yi) is considered in the real image(Cartesian plane), many possible lines are passingthrough this point, and each of these lines maps to adifferent point in the Hough (r–u) space. If the point(xi, yi) is known, although r and u are the unknownvariables in equation (2), the locus of all such linesin the x–y space maps to cisoidal curves in theHough parameter space. This point-to-curve trans-formation is the Hough transformation for straightlines.

If a set of collinear edge point (xi, yi) with commonparameters ro and uo are present (Fig. 15(c)), theneach point maps to a different curve in the Houghspace. However, all curves must intersect at thepoint (ro, uo), as this is their common straight line(Fig. 15(d)).

A typical method for practically counting the inter-section points (ro, uo) is based on a discreteapproach, called ‘voting’. The Hough transform isalgorithmically implemented by quantizing the par-ameter space into accumulator arrays: in this sense,the Hough space is an accumulator space [7].

Consider the four points in the image shown inFig. 16(a) Hough transform maps them into thefour cisoidal curves of Fig. 16(b). The crossingpoints in the Hough plane are numbered and‘voted’. For example, the intersecting point num-bered as c in Fig. 16(b), corresponds to three intersec-tions (three votes), i.e. the points B, C, and D arecollinear in the real image, as in the inverse Houghtransform shown in Fig. 16(c).

In such a way, the Hough transform defines anaccumulator function: the point grouping problemis converted into a peak detection problem and thelocal maxima represent the more voted points, i.e.the lines that more likely belongs to the real image.The main advantages of the Hough transform areits effectiveness of detecting curves of any shape,even non-regular, and its robustness with respectto image gaps and to image noises. As a matter offact, the local maxima in the accumulation functionare poorly affected from gaps and image noises.

A possible algorithm able to detect straight linesinto an image can follow the steps.

1. A binary image is obtained from the real imageusing an edge detector algorithm (e.g. the Cannyalgorithm [8]).An edge detector algorithm produces a featuredescription into a real image. After this prelimi-nary step, the Hough transform identifies the fea-tures (e.g. straight lines) and their quantity within

Fig. 16 Hough transformation: straight lines detection through a voting procedure

Fig. 15 Hough transformation



the edge description. For every pixel belonging tothe edge (x,y), consider the straight line in theHough space. A straight line in the real image ischaracterized by a point (q, r) in the parameters’space and a pencil of lines through (x0, y0) can berepresented by the curve (2) in the Hough space,with the constants x0 e y0, and the parameters q

and r as unknowns. In such a way, the transform-ation maps each point of the real image into aunique curve in the Hough plane q,r, and n-colli-near pixels in the image plane are mapped into adn curves intersecting in a point (q , r), whosecoordinates are the parameters defining thestraight line in the Cartesian plane.After this preliminary step, the algorithm followsas.

2. Define an accumulator matrix such that itselements correspond to small regions of theHough space (quantisation of the region).The problem at this step is the arbitrary choice ofthe rows and columns to use in the matrix, i.e. thequantisation interval in terms of the subdivisionin small regions of the Hough plane.The accumulator matrix associates everyelement P[q, r] to a possible pair (q, r), with q [

½qmin, . . . ,qmax� and r [ ½rmin, . . . , rmax�. Intervallimits must be selected to include all straightlines of interest in the real image.

3. Initialize to zero every element P[q, r] of theaccumulator matrix.

4. For every pixel (x, y) of the edge do

for (q ¼ qmin, q � qmax, qþþ)

r ¼ x cosqþ y sinq

round r to a possible value

P½q,r� ¼ P½q,r� þ 1

end

The problem at this step is to create for each point(x, y) the curve r ¼ r(u) and then to increment theelements (give votes) of the accumulator matrix.

5. Detect the largest elements of the accumulatormatrix P½q, r�, i.e. apply a peak locationalgorithm.

6. Apply the inverse Hough transform convertingback the peaks to straight lines in the real image.



Documents

Hough transform and thermo-vision for monitoring ...webm.dsea.unipi.it/~saniw/public_html/Bari/P_10.pdf · Obtained results are relevant for monitoring the status of the quality transmission