SciELO - Scientific Electronic Library Online

 
vol.19 issue2Physician and family perception of the results of multiple orthopedic surgeries in children with cerebral palsyCardiovascular risk factors in urban and rural communities from Tinaquillo, Venezuela author indexsubject indexarticles search
Home Pagealphabetic serial listing  

Services on Demand

Article

Indicators

Related links

  • On index processCited by Google
  • Have no similar articlesSimilars in SciELO
  • On index processSimilars in Google

Share


Revista de Salud Pública

Print version ISSN 0124-0064

Rev. salud pública vol.19 no.2 Bogotá Mar./Apr. 2017

http://dx.doi.org/10.15446/rsap.v19n2.34043 

Artículos/Investigación

Video-based assessment of pedestrian behavior: Development and testing of methods

Estudio de comportamiento peatonal basado en video: Desarrollo y prueba de los métodos

Lope H. Barrero-Solano1  , Alfonso Sánchez-Pilonieta2  , Alejandro Forero-Guzman3  , Julián A. Quiroga-Sepúlveda4  , Nestor S. Romero-Barreto5  , Francisco C. Calderon-Bocanegra6  , Sarah Felknor7  , Leonardo A. Quintana-Jiménez8 

1LH: Ing. Industrial. M Sc. Ciencias de la Salud Ambiental. Ph. D. Ciencias de la Salud Ambiental. Departamento de Ingeniería Industrial, Pontificia Universidad Javeriana, Bogotá, Colombia. lopehugo@javeriana.edu.co

2AS: Psic. Maestro en Investigación Educativa. Fundación Centro Internacional de Desarrollo Humano y Educación CINDE. Bogotá, Colombia. asanchez@cinde.org.co

3AF: Ing. Electrónico. Maestro en Ingeniería Electrónica. Ph. D (c) Ingeniería. Pontificia Universidad Javeriana, Bogotá, Colombia. alejandro.forero@javeriana.edu.co

4JQ: Ing. Electrónico. Maestro en Ingeniería Electrónica y de Computadores. Ph. D. Informática y Matemáticas. Departamento de Ingeniería Industrial, Pontificia Universidad Javeriana, Bogotá, Colombia. quiroga.j@javeriana.edu.co

5NR: Ing. Electrónico, Pontificia Universidad Javeriana, Bogotá, Colombia. romero-n@javeriana.edu.co

6FC. Ing. Electrónico. Maestro en Ingeniería Electrónica. Ph. D. Ingeniería, Pontificia Universidad Javeriana, Bogotá, Colombia. Calderonf@javeriana.edu.co

7SF: Artes. M. Sc. Ciencias de Desarrollo Organizacional. Ph. D. Salud Pública. University of Texas, Houston. Estados Unidos de América. sarah.a.felknor@uth.tmc.edu

8LQ: Ing. Industrial. Ph. D. Ingeniería Industrial. Departamento de Ingeniería Industrial, Pontificia Universidad Javeriana, Bogotá, Colombia. lquin@javeriana.edu.co

ABSTRACT

Objective

The aim of this paper is to develop a computer algorithm that analyzes pedestrian behavior at an urban site in Bogota, Colombia, considering that the assessment of pedestrian behavior is a road safety priority.

Methods

Pedestrians were video-taped as they crossed a selected road. An algorithm was developed in order to record, from these videos, pedestrian and vehicle positions and speeds. This information made possible the identification of hazardous behaviors, which were compared through visual assessments.

Results

429 pedestrians crossed the selected road at an average distance of 4.5 meters from vehicles that moved at an average speed of 21 km/h. With a maximum difference of 19 % with respect to visual assessments, the algorithm estimated that 58.5 % pedestrians crossed through non-designated locations; 62.2 % crossed near moving vehicles, and that 41.2 % ran while they were crossing the road.

Conclusions

Video-based analysis can be used to assess pedestrians’ behavior. Future research work should focus on improving both the accuracy and the number of safety parameters of the algorithm.

Key words: Videorecording; pedestrian; validity; developing nations; (source: Mesh, NLM).

RESUMEN

Objetivo

La medición del comportamiento de peatones es una prioridad de la seguridad vial. Por lo anterior, se desarrolló un algoritmo para analizar el comportamiento de los peatones en una zona urbana de Bogotá, Colombia.

Métodos

Los peatones fueron filmados mientras cruzaban la calle. Mediante el algoritmo se midieron las posiciones y velocidades de peatones y vehículos en los videos. Se identificaron los comportamientos riesgosos y se compararon visualmente.

Resultados

429 peatones cruzaron la vía a una distancia promedio de 4.5 metros de los vehículos (velocidad promedio 21 km/h). El algoritmo estimó, con una diferencia máxima de 19 % con respecto a lo observado, que 58.5 % de los peatones cruzaron por zonas incorrectas, 62.2 % cruzaron cerca de vehículos en movimiento y 41.2 % corrieron al cruzar.

Conclusiones

El análisis basado en video puede utilizarse para medir el comportamiento de los peatones. Los trabajos sobre el tema que se realicen en el futuro deben enfocarse en mejorar la precisión y los parámetros de seguridad del algoritmo.

Palabras-clave: Grabación de video; peatones; validez de las pruebas; países en desarrollo; (fuente: DeCS, BIEREME).

Assessing road user behavior is essential for the design of safe road environments. Understanding users’ errors, motivations and expectations when using road systems allows for an infrastructure design that minimizes human error, elicits safe behavior and reduces the probability of serious injury 1-5. This paradigm of road environment design is at the heart of road safety philosophies in leader countries such as Sweden (Vision Zero program) and the Netherlands (Sustainable Safety program) 6. The assessment of pedestrian behavior is a priority in today’s global society 7 and, in particular, in developing countries, where they are considered as the most vulnerable road user 8.

There are a variety of ways to assess pedestrian behavior, including the evaluation of pedestrian choices under simulated and real scenarios. Simulated scenarios have been especially valuable in understanding pedestrian behavior 9-14. On the other hand, studies on pedestrian behavior in real scenarios allow researchers to describe and quantify the frequency of common risky behaviors 15-17, as well as testing theories that predict pedestrians’ choices and reactions to traffic conditions.

Video-based methods may constitute a feasible option to assess pedestrians’ road-crossing choices and interactions with vehicles around them. However, most computer applications that focus on evaluating pedestrian behaviors are still at the level of algorithm development for detecting and tracking moving objects 15,18-20 or are not easily available or well suited for the analysis of a variety of traffic conditions 15-17,21-23. Only a few studies have reported the use of flexible applications for road safety with a main focus on assessing the trajectory, traffic volume and/or traffic object speed 19,24,25.

The present study reports the feasibility of a new video-based algorithm to conduct automated analysis of pedestrians’ behavior at a busy urban location in Bogotá, Colombia.

METHODS

Design

In this study, a computer algorithm was developed to gather information on pedestrian behavior from videos taken in a busy urban location in Bogotá, Colombia. Urban locations considered for this study included those with high frequency of pedestrians crossing the road and those were video-camera recording above street level (8th to 10th floor) was possible. To assess the validity of the video-based information, selected behaviors of pedestrians assessed with the algorithm that was created for the video analysis were compared whit those assessed visually.

Data collection procedures

Data collection was done through videotaping pedestrians as they crossed a selected road (Figure 1). An audio video interleave format (AVI) video was taken from a tall building that was adjacent to the road of interest. The view scope of the recordings allowed for the analyses of pedestrian crossings at the corner through designated areas (i.e., the crosswalk or pedestrian crossing) and up to approximately 30 meters from the corner. The video recording angle and distance did not allow to positively identifying specific pedestrians; therefore, it was not possible to describe the demographics of observed pedestrians. Letters requesting permission to place video recording equipment on an upper level of buildings were sent to their administrators. The present study reports the results obtained from videos taken at the corner of 53rd street and 11th avenue in Bogotá, Colombia. Standard SONY camcorder DCR SR82 equipment was used.

Black arrows indicate the pedestrian walking directions of interest; *Traffic lights for vehicles going from north to south on the 13th Avenue

Figure 1 Video recording location 

Algorithm development and testing

The objective of the study was to develop an algorithm that allowed analyzing automatically videos to identify and record pedestrian position and pedestrian distance from motor vehicles, as well as speeds of pedestrians and vehicles during the crossing of a specific road segment. In particular, the study aimed to identify potentially hazardous pedestrian behaviors, i.e., to assess if pedestrians: 1. Crossed in designated areas (i.e., designated locations for pedestrian crossing, commonly known as crosswalks); 2. rossed near moving vehicles that were at a distance less than 2 meters away, and; 3. Ran when crossing a road, which is a measure of conflict between pedestrians and drivers 17.

The recorded video was used to refine a version of an algorithm for movement detection that was previously developed by the authors of this paper. The algorithm was developed using C++ language and uses libraries for image processing from OpenCV (Open computer vision) 26.

The algorithm comprises of a 6-step process: 1) video reading, background modeling using a code book algorithm and extraction of first plane pixels; 2) morphological filtering, object segmentation and indexing; 3) contour extraction for each object in the first plane; 4) calculation of the area moments based on the contour of each object; 5) classification of each object as vehicle, single pedestrian or cluster of pedestrians (to train the classifier, a support vector machine was used), and 6) follow-up 27-29. Based on sizes and shapes of objects, the algorithm classifies a single pedestrian. Bigger objects are classified using a support vector machine that determines whether the object is a vehicle or a cluster of pedestrians. Multiple versions of the algorithm were created in the process of its refinement in order to improve its ability to identify and follow objects correctly in a user-specified area of interest in the video.

The use of the algorithm includes loading the video on the developed application and modifying the algorithm code to indicate the correspondence between pixels in the video and meters on the road, as well as specification of the area of interest. Also, the user must indicate a particular segment within the area of interest. Every person who crosses that segment is counted by the algorithm and a visual count will be shown on the screen. This preparation process takes only a few minutes.

The analyses are executed in real-time as the video progresses. However, the algorithm requires additional calibration from the user as the video progresses, which is achieved by indicating, whenever possible, that there is not a moving object in the area of interest. The algorithm runs on a computer with a standard processor and does not require additional software aside standard computer operating systems.

At the end of the video, the algorithm produces a record of positions (coordinates) and speeds (meters per second) throughout the cross for each object identified in the video at a frequency of 30 Hz. Recorded speeds were averaged over a moving window of 1 second before further analysis. To ascertain whether pedestrians crossed through the crosswalk or not, the coordinates of their specific positions need to be compared with the coordinates defining the crosswalk in the video. To assess whether pedestrians crossed running, average recorded pedestrian speeds need to be dichotomized with an adjustable cutoff that distinguishes between walking and running according to previous reports 13,30-32. Similarly, to evaluate if pedestrians started running at any time during the cross of the road, peak accelerations are estimated from the recorded speeds and dichotomized with various adjustable cutoffs that may indicate significant changes in speed.

To determine the validity of these estimations, the recorded automated algorithm-based pedestrian behaviors were compared with the visual-based assessment of pedestrian behaviors of the same video carried out by a trained analyst. It was possible to achieve at least moderate reproducibility of the analyst’s observations as measured by the Cohen’s Kappa coefficient. Specifically, the following reproducibility values of visual assessment for different behaviors were estimated: pedestrians crossed through the crosswalk (Cohen’s Kappa=95.2); pedestrians crossed at a distance less than 2 meters from a moving vehicle (Cohen’s Kappa=52.1); pedestrians crossed at less than 2 meters from an approaching vehicle (Cohen’s Kappa=54.0), and pedestrians ran while crossing the road (Cohen’s Kappa=43.8). Nevertheless, the analyst showed very poor reproducibility when identifying if pedestrians started crossing at any time.

The comparison between the results obtained in the visual observations and the observations based on the algorithm required for the validity analysis was time-intensive and took up to 3 hours for a 9-minute video. Therefore, a total of 27 minutes of video were randomly selected to perform the analyses. Agreements between manual and automated video-based estimates were measured through percentage of agreement, Spearman’s rank correlations and estimations of sensitivity and specificity. The latter measure is preferred for nominal dichotomous data such as the data compared in this study 33. The Spearman’s rank correlation is a convenient measure that can be easily interpreted but can only be generalized to other populations with similar prevalence in the observed behaviors 33. Lastly, although the Cohen’s Kappa coefficient is an appropriate measure of reproducibility 34, it has also been estimated because it has been frequently used in the literature as a measure of validity.

RESULTS

Based on visual counting, a total of 449 persons crossed the roads during the 27 minutes of video that were analyzed in the present study. The algorithm identified 165 objects of the pedestrian type. The visual assessment showed that the algorithm identified pedestrian objects with up to 7 pedestrians in each object; and most frequently the algorithm made objects containing two persons (34 %). Based on the size of the object (area measured in pixels of the video), the algorithm estimated that 429 pedestrians crossed the segment of interest of the road. Twenty (4.5 %) pedestrians that crossed the segment of interest of the road were missed by the algorithm.

Pedestrians who were missed by the algorithm were not related to whether they crossed through designated locations (Chi-squared test, p value=0.1170) or crossed at less than 2 meters from moving vehicles (Chi-squared test, p value=0.3133); however, the estimated proportion of pedestrians running when crossing was higher among the ones who were detected by the algorithm (46.9 %) than among those who were not detected (20 %) (Chi-squared test, p-value=0.0184), and the estimated proportion of pedestrians running at any time during the cross was higher among those who were detected by the algorithm (71.8 %) than among those who were not detected (30 %) (Chi-squared test, p- value<0.0001).

The algorithm successfully followed the detected pedestrians throughout the cross from sidewalk to sidewalk in the specified area of interest. The expected error in the estimations of distance between the center of different objects of interest (i.e., pedestrian and vehicles) was 0.3 meters. Assuming a negligible error in time estimation, the average speed of pedestrians during the crossing and the instantaneous speed of the closest vehicle were also estimated with a maximum error of 7 meters per hour (Table 1).

Table 1 Recorded pedestrian-vehicle distance and vehicle speeds by size of the pedestrian object 

*The number of pedestrians in each pedestrian object detected by the algorithm was counted based on manual observations

Among those pedestrians that were detected by the algorithm, the agreement between visual and automated observations was at least moderate depending on the assessed behavior and the measure of agreement that was used (Table 2). In general, the percentage of agreement and the sensitivity and specificity measures showed good agreement between both methods. In contrast, for the observation of whether pedestrians cross at 2 meters or less from a moving approaching vehicle or not and the observation of whether pedestrians cross running at any time during the cross or not (for example if the pedestrian starts the cross walking and then starts running), low correlations were found (28.7-37.4), as well as only fair agreement by using Cohen’s Kappa according to the criteria proposed by Landis & Koch 35.

Table 2 Agreement between manual and algorithm-based (automated) observation of pedestrians’ behaviors 

Rs: Spearman Correlation; *The visual observations could identify whether the vehicle that was at 2 meters away or closer was approaching to the pedestrian. The algorithm-based (automated) observation could not identify whether the vehicle was approaching the pedestrian or not; **This was the largest observed agreement for different cutoffs for running (mean pedestrian crossing above 1.9 meters/s) ; ***This was the largest observed agreement for different cutoffs for speed change (peak acceleration above 0.6 meters/s2).

These levels of agreement resulted in group-based differences between automated and visual-based estimations of the prevalence of different pedestrian behaviors that range from -6.4 % to 19.3 % depending on the behavior parameter that is assessed (Table 3).

Table 3 Group-based differences in estimated pedestrian behaviors between manual and automated observations 

(1) Estimations based on manual observations including all pedestrians; (2) Estimations based on manual Observations including only pedestrians detected by the algorithm

DISCUSSION

In this study an automated video-based method for researching pedestrian behavior was developed and tested. The application is deemed inexpensive since it does not require especial hardware and is based on an algorithm code that is freely available to be used in other scenarios, as well as to be potentially improved by other researchers. This accessibility characteristic is important as algorithms for road safety applications are not commonly available and the needs to assess road safety conditions are increasing 3.

The application developed in this research measured validly the proportion of pedestrians crossing through designated locations and the number and proportion of pedestrians who crossed it running. However, the application showed moderate disagreement with manual observations regarding the proportion of persons who crossed the segment of interest at less than 2 meters from an approaching vehicle, as well as the proportion of people who run at any time during the cross of the road.

The size of the differences found between algorithm-based and visual-based observations of the prevalence of various pedestrian behaviors was deemed acceptable. As the algorithm did not detect 4.5% of the pedestrians crossing the segment of interest of the road during the period of assessment, all behaviors that were evaluated were affected. However, this loss of information resulted in differences of only 0.3 to 1.9 % in the overall estimations of the prevalence of the assessed behaviors (Table 3). Regarding the estimation of the proportion of pedestrians crossing through designated places, the results obtained through the algorithm had a maximum difference of 6.4 % in comparison with those obtained in the visual assessment. Such difference cannot be easily compared to differences reported in previous studies as traffic conditions in this study may differ from those in other studies. Nevertheless, a rough comparison with one paper reporting this difference indicates that the error size found in this work is consistent with such results 19.

Differences between automated and visual-based estimations regarding the proportion of persons who crossed running and the proportion of people who crossed near moving vehicles were moderate and were affected by the specific behavior definition used and by the reproducibility of the visual method that was used as reference method. When the proportion of persons crossing near any moving vehicle was estimated the algorithm showed differences below 3.8 % in relation to visual-based assessments (Table 3). However, if the condition that only moving approaching vehicles are considered in the estimation is taken into account, this difference increases to 19.3 %, which is explained as the algorithm was not designed to distinguish vehicles moving near pedestrians from vehicles moving near pedestrians and also approaching them. The function to make such distinction should be a matter of future developments of this algorithm.

In addition, the algorithm did not show important differences with visual observations regarding the detection of pedestrians who ran during their whole cross (up to -5.7 %), however if only pedestrians who run at any time during the cross are considered, these differences are up to 16.8 %. The latter estimation may be more relevant as a measure of road safety because it may be better related to the occurrence of pedestrian-vehicle conflicts in roads 16, in spite of that, running may be in fact a protective behavior (e.g., potential reactions of pedestrians to avoid being struck by vehicles).

The disagreement between automated and visual assessments to estimate the prevalence of pedestrians who run any time during the cross can be attributed to the analyst difficulty to observe this behavior. Thus, in the case of this behavior the automated analysis based on the assessment of speed change may be a more accurate method to investigate whether pedestrians run at any time during the cross or not.

This study had several limitations. First, there is uncertainty regarding the performance of the algorithm in different locations. However, it is foreseeable that the algorithm may result in more accurate estimates in less busy locations since objects (i.e., pedestrians and vehicles) that are more separated may be easier to identify and track down. Furthermore, different locations may result in different sources of occlusion and different recording angles that may affect the results to be obtained.

On the other hand, the algorithm is not yet suited for night or rainy conditions. Finding the solution to these potential shortcomings related to pedestrian tracking in crowded areas is a complex task and is the focus of attention of recent works on human detection 36. Addressing these problems will be a focus of future enhancements to the algorithm conducted by the authors of this paper.

It can be concluded that video-based automated analysis of road safety conditions are feasible and can be successfully used to assess pedestrian behavior in urban scenarios. Future work on this tool shall include the creation of prediction models to estimate variables such as “post-encroachment time”, “time to accident”, “compromised pedestrian crossings” 37 and “vehicle deceleration rate” 16).

REFERENCES

1. Rumar K. Transport safety visions, targets and strategies: beyond 2000. 1st European Transport Safety Lecture. Brussels: European Transport Safety Council; 1999. [ Links ]

2. World Health Organization. Global status report on road safety: supporting a decade of action. Geneve: World Health Organization; 2013. [ Links ]

3. Hakkert AS, Gitelman V, Vis MA. Road safety performance indicators: Theory. Deliverable D3, 6. 2007. [ Links ]

4. Räsänen M, Lajunen T, Alticafarbay F, Aydin C. Pedestrian self-reports of factors influencing the use of pedestrian bridges. Accid Anal Prev. 2007;39(5):969-73. Doi: 10.1016/j.aap.2007.01.004. [ Links ]

5. Bliss T, Breen J. Country guidelines for the conduct of road safety management capacity reviews and the specification of lead agency reforms, investment strategies and safe system projects. Washington D.C: World Bank Global Road Safety Facility; 2009. [ Links ]

6. Koornstra M, Lynam D, Nilsson G. SUNflower: a comparative study of the development of road. Leidschendam: SWOV Institute for Road Safety Research; 2002. [ Links ]

7. Lee C, Abdel-Aty M. Comprehensive analysis of vehicle-pedestrian crashes at intersections in Florida. Accid Anal Prev. 2005;37(4):775-86. doi: 10.1016/j.aap.2005.03.019. [ Links ]

8. World Health Organization. Global status report on road safety: time for action. Geneve: World Health Organization; 2009. [ Links ]

9. Schiff W, Oldak R, Shah V. Aging persons' estimates of vehicular motion. Psychol Aging. 1992;7(4):518-25. doi: 10.1037//0882-7974.7.4.518. [ Links ]

10. Moyano-Díaz E. Teoría del comportamiento planificado e intención de infringir normas de tránsito en peatones. Estud. psicol. (Natal). 1997;2(2):335-48. doi: 10.1590/s1413-294x1997000200008. [ Links ]

11. Evans D, Norman P. Understanding pedestrians' road crossing decisions: an application of the theory of planned behavior. Health Educ Res. 1998;13(4):481-9. doi: 10.1093/her/13.4.481-a. [ Links ]

12. Pitcairn TK, Edlmann T. Individual differences in road crossing ability in young children and adults. Br J Psychol. 2000;91(Pt 3):391-410. 10.1348/000712600161899. [ Links ]

13. Oxley JA, Ihsen E, Fildes BN, Charloton JL, Day RH. Crossing roads safely: An experimental study of age differences in gap selection by pedestrians. Accid Anal Prev. 2005;37(5):962-71. doi: 10.1016/j.aap.2005.04.017. [ Links ]

14. Holland C, Hill R. The effect of age, gender and driver status on pedestrian’s intention to cross the road in risky situations. Accid Anal Prev. 2007;39(2):224-37. doi: 10.1016/j.aap.2006.07.003. [ Links ]

15. Veeraraghavan H, Masoud O, Papanikolopoulos NP. Computer vision algorithms for intersection monitoring. IEEE Transactions on Intelligent Transportation Systems. 2003;4(2):78-89. doi: 10.1109/tits.2003.821212. [ Links ]

16. Malkhama S, Tight M, Montgomery F. The development of an automatic method of safety monitoring at Pelican crossings. Accid Anal Prev. 2005;37(5):938-46. doi: 10.1016/j.aap.2005.04.012. [ Links ]

17. Hubbard SM, Bullock DM, Mannering FL. Right turns on green and pedestrian level of service: Statistical assessment. Journal of Transportation Engineering. 2009;135(4):153-9. doi: 10.1061/(asce)0733-947x(2009)135:4(153). [ Links ]

18. Curio C, Edelbrunner J, Kalinke T, Tzomakas C, Seelen W. Walking pedestrian recognition. IEEE Transactions on Intelligent Transportation Systems. 2000;1(3):155-63. doi:10.1109/6979.892152. [ Links ]

19. Somasundaram G, Morellas V, Papanikolopoulos N. Counting pedestrians and bicycles in traffic scenes. In: 12th International IEEE Conference on Intelligent Transportation Systems. St. Louis: IEEE Transactions on Intelligent Transportation Systems; 2009. [ Links ]

20. Wang H, Lu R, Wu X, Zhang L, Shen J. Pedestrian detection and tracking algorithm design in transportation video monitoring system. In: International Conference on Information Technology and Computer Science. Beijing; 2009. [ Links ]

21. Lord D. Analysis of pedestrian conflicts with left-turning traffic. Transportation Research Record: Journal of the Transportation Research Board. 1996;1538:61-7. doi: 10.3141/1538-08. [ Links ]

22. Houten, R, Retting R, Farmer C, Houten J. Field evaluation of a leading pedestrian interval signal phase at three urban intersections. Transportation Research Board. 2000;1734:86-92. doi: 10.3141/1734-13. [ Links ]

23. Hamed MM. Analysis of pedestrian's behavior at pedestrian crossings. Safety Science. 2001;38(1):63-82. doi: 10.1016/S0925-7535(00)00058-8. [ Links ]

24. Kumar P, Ranganath S, Weimin H, Sengupta K. Framework for real-time behavior interpretation from traffic video. IEEE Transactions on Intelligent Transportation Systems. 2005;6(1):43-53. doi: 10.1109/tits.2004.838219. [ Links ]

25. Park S, Trivedi MM. Analysis and query of person-vehicle interactions in homography domain. In: 4th ACM international workshop on Video surveillance and sensor networks. New York: ACM Press; 2006. [ Links ]

26. Bradski G, Kaehler A. Learning OpenCV: Computer Vision with the OpenCV Library. 1st ed. Sebastopol: O’Reilly; 2009. [ Links ]

27. Forero A, Parra C. Implementación de una herramienta para la adquisición de variables de tráfico: peatones y vehículos, utilizada en la evaluación del riesgo de peatones que cruzan la vía. In: XIII Simposio de tratamiento de señales, imágenes y visión artificial. Bucaramanga; 2008. [ Links ]

28. García, C, Romero N, Quiroga J. Detección y seguimiento de personas en un cruce peatonal. In: XIV Simposio de Tratamiento de Señales, Imágenes y Visión Artificial. Pereira; 2009. [ Links ]

29. Urrego GE, Calderón FC, Forero A, Quiroga JA. Adquisición de variables de tráfico vehicular usando visión por computador. Revista de Ingeniería. Universidad de los Andes. 2009;30:7-15. doi: 10.16924%2Friua.v0i30.222. [ Links ]

30. Wakim CF, Capperon S, Oksman J. A Markovian model of pedestrian behavior. In: 2004 IEEE International Conference on Systems, Man and Cybernetics. Netherlands; 2004. [ Links ]

31. Te Velde AF, Kamp J, Barela JA, Savelsbergh GJP. Visual timing and adaptative behavior in a road-crossing simulation study. Accid Anal Prev. 2005; 37(3):399-406. doi: 10.1016/j.aap.2004.12.002. [ Links ]

32. Zhang R, Li Z, Hong J, Han D, Zhao L. Research on Characteristics of Pedestrian Traffic and Simulation in the Underground Transfer Hub in Beijing. In: Fourth International Conference on Computer Sciences and Convergence Information Technology. Seoul: IEEE.; 2009. [ Links ]

33. Maclure M, Willett WC. Misinterpretation and misuse of the kappa statistic. Am J Epidemiol. 1987;126(2):161-9. doi: 10.1093/aje/126.2.161. [ Links ]

34. Guggenmoos-Holzman I. The meaning of Kappa: Probabilistic concepts of reliability and validity revisited. J Clin Epidemiol. 1996;49(7):775-82. doi: 10.1016/0895-4356(96)00011-x. [ Links ]

35. Landis JR, Koch GG. The measurement of observer agreement for categorical data. Biometrics. 1977;33(2):159-74. doi: 10.2307/2529310. [ Links ]

36. Fehr D, Sivalingman R, Morellas V, Papanikolopoulos N, Lotfallah O, Park Y. Counting people in groups. In: Sixth IEEE International Conference on Advanced Video and Signal Based Surveillance. Genoa: IEEE; 2009. [ Links ]

37. Hubbard SM, Awwad R, Bullock D. Assessing the impact of Turning Vehicles on Pedestrian Level of Service at Signalized Intersections: A New Perspective. Transportation Research Board. 2007;2027:27-36. [ Links ]

Received: October 18, 2015; Revised: August 10, 2016; Accepted: December 08, 2016

Grant sponsor :

National Institute of Health Fogarty International Center; Grant number: 5 D43 TW007564

Creative Commons License This is an open-access article distributed under the terms of the Creative Commons Attribution License