ce. September 1-3. 2010 In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds), 1APRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3. 2010
Ю08. Building bound-
i and lidar data. Inter-
Remote Sensing and
^p. 693-698.
IMPROVEMENT OF A PROCEDURE FOR VEHICLE DETECTION AND TRACKING
BY BASE FRAME UPDATING AND KALMAN FILTER
G.Artese 3
). Extraction of build-
mbination of various
letry, Remote Sensing
W4), pp. 71-76.
;i Ideation accuracy as-
80(1). pp. 185-201.
of lidar and imagery
mmetric Engineering
'lass-guided building
mmetric Engineering
rtion from aerial im-
'omputer Vision and
Kubik, K., 2005. Us-
ion of lidar data and
. Information Fusion
id Kubik, K., 2007.
ser scanner data and
ation and sensitivity
y and Remote Sens-
9., 2009. A compari-
action from airborne
)ics in Applied Earth
11-20.
iing extraction from
Engineering 131(1),
1 analysis of monoc-
IEEE Trans, on Pat-
). pp. 311-326.
n of high-resolution
building extraction,
note Sensing 62(1),
nent of comprehen-
ng footprint extrac-
id Remote Sensing
i. Symmetric stereo
. IEEE Conference
, Voi. 2, San Diego,
). Multi-scale solu-
nage data. Interna-
md Geoinformation
ding edge detection
emational Archives
Spatial Information
‘’Land Planning Dept., University of Calabria, 87036 Rende, Italy - g.artese@unical.it
Commission III - WG III/5
KEY WORDS: Digital Photogrammetry, Vehicle Tracking, Kalman Filter, Change Detection
ABSTRACT:
For the vehicle detection and tracking in the roundabouts, several difficulties have to be faced, e.g. the use of non nadiral
perspectives, vehicle images overlapping and radiometric differences between top and side of the cars: the silhouette of the cars is,
besides, not regular and the size is variable. In the paper, the upgrade of a procedure for detecting and tracking vehicles in a
roundabout is presented. The image subtraction technique is used, along with geometric and radiometric filters. The problems due to
the variation of the background radiometric characteristics are discussed. Kalman filter has been used to foreseen the position of
vehicles for a better tracking, and to improve the determination of the trajectories. The results of a test are presented.
1. INTRODUCTION
Object detection and tracking can be very' useful in a range of
industries and applications, like surveillance, vehicle and
pedestrian tracking. Trajectory and velocity estimation of
vehicles is very important for the management of traffic, above
all in the urban zones and for road intersections.
The most used methods for traffic monitoring allow to obtain
the number of vehicles crossing a given section of a road;
automatic sensors (pressure hose sensors, magnetic buried
loops) have been used for decades, to automatically count the
axes or the vehicles, but no information about trajectories are
generally obtained.
Several procedures, based on digital photogrammetry
techniques and computer vision have been proposed in the last
years for the estimation of traffic flow. New available
technologies have been also used, and multisensor systems have
been tested (Grejner-Brzezinska et al„ 2007, Toth and Grejner-
Brzezinska, 2007. Yao et al„ 2009, Goyat et al., 2009). The
main advantage of the computer vision techniques is the
possibility to obtain information regarding not only the number
of entering and outgoing vehicles, but also their trajectories.
Some authors proposed algorithms set up by using a single
camera (Reulke et al., 2002, Broggi and Dickmanns, 2000) or
multi-camera systems (Pedersini et al., 2001). In more complex
systems, images obtained with infrared cameras are also used
(Dalaff et al., 2003, Hinz and Stilla, 2006, Kirchhof and Stilla,
2006). The detection of vehicles is generally performed by
using procedures based on the segmentation of the groups of
pixels having similar chromatic values, and on the edges
extraction. Morphologic operations (regions filling and erosion,
edges dilatation) can be performed to optimize the results.
Depending on the kind of sensor data, a data base with the
characteristics of the vehicles can be used for facilitating the
recognition, while geometric conditions, such as the parallelism
between street edge and vehicle trajectories can be imposed
only for straight roads (Puntavungkour and Shibasaki. 2003).
If the differences of images are used, the frames obtained by a
camera are compared to detect the presence and the position of
the vehicles. One can follow essentially two ways: to compare
every frame with the previous one, or to compare every frame
with a base frame (background) without vehicles.
Due to the short time interval between two consecutive frames,
the background is practically identical in the first case, and it is
possible to detect the vehicles in motion, but not those
immobile ones; if this way is followed, the block-based motion
estimation is usually chosen, by using high performance
computers (Min Tan and Siegel, 1999). Correlation techniques
are used, by subdividing two consecutive frames in small
blocks, and by obtaining for every block the movement vector:
a neighbourhood window of pixels in a given image, centred on
a specific pixel, is searched over a larger neighbourhood
window of pixels in the previous image, centred on the same
pixel. It is assumed that from frame-to-frame in a video
sequence pixel intensity values do not change and that the video
source is stationary. The position where the minimum absolute
or squared difference is obtained, gives the motion vector for
the centre pixel. To save run time, the block-based motion
estimation is run only for a subset of pixels in the image and the
results are interpolated over the entire image.
If a base frame is used, it is possible to detect both mobile or
immobile vehicles, but some problems must be solved,
essentially due to:
- variations of the lighting conditions, that make practically
impossible to use a single base frame:
- presence of noises in the images;
- shadows projected from the vehicles, trees and surrounding
buildings;
- movements of the video camera (oscillations and spins, even
if modest, make the elimination of the background difficult and
create false moving objects).
Other difficulties to overcome are the superimposition of the
images of close vehicles and the necessity of a real time
elaboration.
A key element for many target tracking algorithms is an
accurate background subtraction (Seki and Wada, 2003,
Isenegger et al., 2005). An overview of several techniques for
background detection and updating has been made by Cheung
and Kamath (2004). In this paper, simple Frame Difference
(FD), Median Filter (MF) (Cucchiara at al., 2003),
Approximated Median Filter (AMF) (McFarlane and Schofield,
1995) , Kalman Filter (KF) and Mixture of Gaussian (MoG)
(Friedman and Russell, 1997, Stauffer and Grimson, 1999) are
compared. The best results are obtained by MoG and MF,
followed by AMF. In spite of its simple implementation, low
storage requirements and computing rapidity, Approximated
Median Filter shows good performances. Also the Running
Gaussian Average (RGA) (Wren et al., 1997) can be used with
good results. More sophisticated approaches, like Bayesian
Background Estimation (Rahimizadeh et al., 2009) or
multimodal information integration (Kato and Wada, 2004) has
been proposed. Kalman Filter for background detection has
been implemented in commercial software packages (MVTec ,
2009).
In the following, a technique for detecting and tracking vehicles
in a roundabout is described. Image subtraction, geometric and
55
