Collecting pedestrian trajectories
Introduction
The dynamic of large pedestrian streams has a wide range of application ranging from animation of crowds in movies [1], planning and optimisation of transport infrastructures or the design of pedestrian facilities in buildings and at mass events [2]. While qualitative correctness is sufficient for movies, quantitative correctness is necessary for the other applications. In particular, in application areas dealing with the safety of people a reliable experimental database is mandatory. The data provide a basis for quantifications in legal regulations, guidelines and manuals for the design of pedestrian facilities or for the validation of computer models to simulate the movement of crowds dynamically.
However, until now the experimental database is insufficient and even the simplest questions, such as the increase in the capacity of bottlenecks with the width or the density in a crowd where the formation of jams are expected, are up to now discussed controversially [3]. To understand and to model pedestrian dynamics, reliable empirical data are needed. In recent years, we have performed an extensive series of well defined experiments to study the movement of pedestrians in different situations.
We are developing a software to automatically extract trajectories from video recordings of movements on plane ground [4] and uneven terrain [5]. The program is able to handle lens distortion and high pedestrian density. For spatial trajectories, stereo recordings are needed. To detect pedestrians without markers, we newly introduced a method based on the analysis of the height field of stereo recordings.
Collecting the exact trajectories of every person allows a detailed analysis of movement and verification of microscopic models in space and time [6]. Microscopic modelling needs microscopic behaviour trajectories [7]. The set of trajectories of all pedestrians provides data like velocity, flow, density and individual distances at any time and position, thus lane formation and local densities can be analyzed [8]. In [9], we showed that the flow through a bottleneck increases continuously with the width. How precise trajectories could help to understand the formation of jams in pedestrian streams could be found in [10].
In our experience, some of the contradictions in the literature can be traced back to insufficient methods of data capturing or inadequate resolution of the measurement in time and space [11]. To improve this situation, we determine the trajectories as accurately as possible and develop measurement methods providing high resolution of quantities like density, flow or velocity in combination with small scatter [12].
This paper gives a survey of related work for detecting people with camera techniques in Section 2. Section 3 summarises the experiments we have accomplished and Section 4 describes the different strategies to extract the trajectory of every person out of the video recordings. Within these strategies we focus on the markerless detection method and show some results in Section 5.
Section snippets
Related work
Techniques for trajectory extraction without markers for single pedestrians in crowds using monocular cameras are not as robust as techniques using stereo cameras. Papers like [13], [14], [15], [16], [17] all report a false detection rate of more than 10%. Typically the decrease in the false detection rate induces the increase in false positive detections. For our purpose, this is not acceptable, because we need nearly no error to get reliable data for further analysis.
Detection techniques,
Experiments
Performing experiments under laboratory conditions gives the opportunity to analyse parameters of interest under well defined constant conditions. The variability allows a survey of a parameter range e.g. for the bottleneck width or length, or the density inside a corridor. Parameters can be set to values seldom seen in field studies (e.g. very high densities).
For self-initiated experiments, the location and the structure of the test persons (e.g. culture, fitness, age, gender, size) can be
Detection and tracking
The overall goal of the extraction process are trajectories as exact as possible from every person at any time. For this reason, markers are used for easier detection. This improves the robustness of the automatic extraction. Nevertheless in Section 4.3 the newly introduced algorithm for the markerless detection is also discussed.
For the same reason of exactness, all automatic results are inspected by humans who are able to correct the trajectories directly within our software. Under laboratory
Results
One result of the markerless detection can be seen in Fig. 6. The pyramids on the left are coloured according to their level inside the pyramid. The top most pyramid has a green, the second a red and the third a blue colour. The blue one locates roughly the transition from head to shoulders. The latter ellipses can cover more than one person. For the visualised frame, this induces to three persons without a corresponding pyramid. This results from the deletion of small pyramids inside big
Conclusion and outlook
In this paper, we gave an overview of our experiments of people movement and the different strategies to collect precise trajectories out of overhead video recordings from these experiments. The recent approach for the markerless detection is discussed in more detail. Using directly the perspective height field from stereo recordings is a fast method for the perpendicular view. All introduced techniques allow an automatic extraction of all trajectories with small error, which we need especially
Acknowledgements
This study was performed within the project funded by the German Research Foundation (DFG) KL 1873/1-1 and SE 1789/1-1 and the project Hermes funded by the Federal Ministry of Education and Research (BMBF) Program on “Research for Civil Security—Protecting and Saving Human Life”.
Maik Boltes completed his apprenticeship as Mathematical–Technical Assistant at the Forschungszentrum Jülich in 1994. From 1994 until today he works in the field of computer graphics at the Forschungszentrum Jülich. Beside this he studied Mathematics and Informatics at the FernUniversität Hagen and received the diploma with distinction in 2005 with a thesis on adaptive mesh simplification techniques for boundary conditions in virtual reality environments and got therefore the young scientist
References (38)
- et al.
Autonomous pedestrians
Graphical Models
(2007) - et al.
Methods for measuring pedestrian density, flow, speed and direction with minimal scatter
Physica A
(2010) - et al.
Pedestrian detection and tracking at crossroads
Pattern Recognition
(2004) - et al.
People detection and tracking using stereo vision and color
Image Vis. Comput.
(2007) Stereo person tracking with adaptive plan-view templates of height and occupancy statistics
Image Vis. Comput.
(2004)- et al.
Robust pedestrian detection and tracking in crowded scenes
Image Vis. Comput.
(2009) - A. Schadschneider, W. Klingsch, H. Kluepfel, T. Kretz, C. Rogsch, A. Seyfried, Encyclopedia of Complexity and System...
- et al.
Automatic extraction of pedestrian trajectories from video recordings
- et al.
Using stereo recordings to extract pedestrian trajectories automatically in space
Improved methods for checking forces based models of pedestrian dynamics
Generalized centrifugal force model for pedestrian dynamics
Phys. Rev. E
Analysis of bottleneck motion using Voronoi diagrams
New insights into pedestrian flow through bottlenecks
Transp. Sci.
Phase coexistence in congested states of pedestrian dynamics
Enhanced empirical data for the fundamental diagram and the flow through bottlenecks
Cited by (0)
Maik Boltes completed his apprenticeship as Mathematical–Technical Assistant at the Forschungszentrum Jülich in 1994. From 1994 until today he works in the field of computer graphics at the Forschungszentrum Jülich. Beside this he studied Mathematics and Informatics at the FernUniversität Hagen and received the diploma with distinction in 2005 with a thesis on adaptive mesh simplification techniques for boundary conditions in virtual reality environments and got therefore the young scientist award from the FernUniversität Hagen. Since 2009 he is Ph.D. student in Mathematics at the University of Cologne and works on the automatic extraction of pedestrian trajectories in crowded scenes.
Armin Seyfried studied Theoretical Physics at the Bergische Universität Wuppertal from 1988 to 1996. In the course of his diploma project and Ph.D. thesis, which he finished in 1998, he focused on many particle systems, high energy physics and parallel computing. After his Ph.D. he specialised in the fire safety field. Since 2004 he has been establishing a new research group for pedestrian dynamics and fire simulations at the Jülich Supercomputing Centre of the Forschungszentrum Jülich. In 2010, he became Professor for computer simulations for fire safety and pedestrian traffic at the Bergische Universität Wuppertal.