Docstoc

moose is loosedvi

Document Sample
moose is loosedvi Powered By Docstoc
					   A framework for moose early warning
         driver assistance systems
                    Technical report




       ¨
PETER HANDEL, YI YAO, NINA UNKURI AND ISAAC
                   SKOG




          Stockholm Sweden December 2006


IR-EE-SB 2006:058
                                                                                                                           1



   A framework for moose early warning driver
                                      assistance systems
                             Peter Händel, Yi Yao, Nina Unkuri and Isaac Skog




                                                         Abstract

          Encounters between big game, such as the moose, and private cars often result in severe injuries and
     death. A vehicle-based moose detection early warning system is a technical countermeasure to increase the
     traffic safety by alerting the driver in case of danger. Based on available off-the-shelf consumer electronics,
     a far infrared night vision system was built to serve as a platform for experiments with moose early
     warning driver assistance systems. Aspects of a moose thermal replica employing a horse are discussed
     and typical vehicle-moose movements are directed and recorded to form a publicly available database
     for research and education. Experiments with in-house developed real-time moose detection software are
     included. Practical considerations, which must be taken into account before employing these kinds of
     advanced driver assistance systems in large scale, are reported as well, based on the long-term use of
     the system.


                                                       Index Terms

          Collision detection, night vision system, advanced driver assistance systems, moose thermal replica,
     Elchtest, early warning system.




  The authors are with the Signal Processing Lab, Royal Institute of Technology, Stockholm, Sweden.
  This work has in part been funded by Skyltfonden under contract EK 50 A 2006:4636. Part of Händel’s work has been carried
out at a sabbatical leave at the Institute of Signal Processing, Tampere University of Technology, Finland. The corresponding
author’s email: ph@kth.se.


December 18, 2006                                                                                                     DRAFT
                                                                                                                              2




Fig. 1.   Typical outcome of an encounter with a Swedish moose. The roof at the A-frame has been exposed by a significant
down-force by the trunk of the moose. The front of the private car is rather unaffected by the encounter. The moose died. Photo:
S. Sundkvist, courtesy of Älgskadefondföreningen.




                                                    I. I NTRODUCTION

A. Car accidents involving moose

   Encounters between wild animals and private cars may result in severe injuries and death. Sweden is a
Nordic country where roughly 50 % of all car accidents are due to situations involving wild animals. For
example, in 2005 there were 33,452 reported accidents with bears, wolves, wolverines, lynxes, moose,
deer, roe deer, otters, wild boars, mouflons, or eagles, but the actual number of wildlife accidents is
estimated to be substantially larger. Typically, a small proportion (3–4 %) of these wildlife accidents
results in injured people. Deer are involved in a majority of these accidents, but the Swedish moose (or
elk, that is, Alces alces) is involved in some 13 % of the accidents. In general, it is relatively harmless
to collide with a wild animal in Sweden, with the exception of the Swedish moose. The latter type of
collision most often results in severe injuries due to the long legs of the moose and its heavy mass, as
illustrated by Fig. 1. Annually, about ten people are mortally injured in this type of encounter, out of
some 500 mortally injured in total. The situation in Sweden as reported above is just an example of a
country where incidents between private cars and big game result in severe injuries and death.

December 18, 2006                                                                                                       DRAFT
                                                                                                                                    3


                                                 700
                                                           2003               2004               2005             2006
                                                 600

                                                 500




                            Reported accidents
                                                 400

                                                 300

                                                 200

                                                 100

                                                  0
                                                  Jan April July Oct Jan April July Oct Jan April July Oct Jan April July Oct
                                                                                     Month




Fig. 2.   Statistics from the police authority STORM database showing the number of reported car accidents with moose in
Sweden for the time period of January 2003 to October 2006. The yearly peaks in October are marked with ’o’.




   There are some 200,000-250,000 moose in Sweden during the winter, with some 100,000 new born at
late spring. Then, at the autumn the population is reduced by the annual moose hunt. Accordingly, the
moose population is regulated and kept at a constant level from one year to another — the number of
car accidents with moose involved is thus expected to remain at a constant level if no novel measures
are taken to decrease the number. The number of accidents involving moose not only is a function of
the actual population but also depends on the time of year, as shown in Fig. 2.


B. Counter measures to increase traffic safety

   Driver-oriented measures to decrease the number of accidents with game include the use of stationary
game danger signs. However, this kind of sign has proved to be ineffective if used alone [1]. This is
simply due to the fact that the probability of encountering game is too low for the driver to take the
signal seriously, even though the probability is orders of magnitude higher than in low risk areas. Other
infrastructural means to prevent accidents include fences, over- and underpasses for animals, odorous
substances and reflective apparatuses [1].
   A motivation for this work is the rapid technical development within advanced driver assistance systems.
Night vision systems have been around for some time; for example, the after-market far infrared (FIR)
Nightdriver system for the Hummer H1 and H2 was launched in 2003 [2]. More recently, original
equipment manufacturer systems have also been available as options when ordering a car; for example,
the near infrared (NIR) Night View system by Mercedes-Benz for their S-class, and the FLIR PathFindIR
far infrared system for the BMW. Also, other vendors provide similar systems for their top-of-the-


December 18, 2006                                                                                                               DRAFT
                                                                                                          4



line models. In the Swedish market, the FIR Hummer system is around 5800 EUR to end consumers,
whereas the pricing for the Mercedes NIR system is listed as EUR 1700 (October 2006). Pricing is
thus still an obstacle for the end consumer that restricts the use of this type of equipment in large
scale. Another obstacle for its use in large scale may be governmental export restriction and licensing,
including constrained terms of use. Due to Moore’s law, a future price reduction is foreseen. In addition,
an increased accessibility for this kind of vehicle accessory is expected — one may note that an export
version has been announced of the FLIR PathFindIR infrared camera that does not require a US export
license.


C. Night vision technology

  The night vision systems are based on infrared technology, in which the observed temperature differ-
ences are converted into a video signal. Besides its basic use as a driver aid with the raw video showing
the heat signatures displayed as a black-and-white image on a head-up display or in the instrument cluster,
video processing may be employed for identification and tracking of objects.
  In the literature, most of the reported work in identification and tracking of objects by infrared vision
is on pedestrian early warning systems, although a simulated software-based moose early warning system
has been studied from a user point of view in [3].
  In [4], shape-independent pedestrian detection is considered by far-infrared images. In [5], [6], methods
are presented based on localization of shape-dependent objects. The work in [4], [5], [6] is based on
single camera far infrared vision. Stereo night vision systems for the purpose of detection is considered
in [7], [8], [9]. System perspectives of night vision systems are considered in [10], [11]. Work on far
infrared pedestrian detection, including fusion from auxiliary sensors, has been reported as well. In [12],
a laser scanner and ego motion sensors are added to a single camera system, and in [13], visual stereo
vision is used in combination with a stereo infrared system. A system based on near infrared images is
presented in [14].
  As reported above, there has been a significant amount of activity within the realm of pedestrian early
warning systems. To the authors’ knowledge, there is basically no published work on vehicle-based early
warning systems for big game such as the moose. However, one may note that the night vision system
used in our study (the Nightdriver-system, for which technical specifications are summarized in Table
I) is commercially used for driver’s manual moose spotting by drivers of the truck fleet of Sysco Food
Services of Atlantic Canada [15]. Typically, each of those drivers spots between four and six moose a
night [15]. Thus, work on automatic moose spotting is indeed motivated from a traffic safety point of

December 18, 2006                                                                                    DRAFT
                                                                                                                 5


                                                      TABLE I

T ECHNICAL SPECIFICATION OF THE EMPLOYED NIGHT VISION UNCOOLED BST (160 × 120)            PIXEL ARRAY FAR INFRARED

                                                 SYSTEM       [2].




                             Spectral response         7-14 µm
                             NTSC video update rate    30 Hz
                             Time to operation         45 s
                             Field of view             10o (horizontal) × 4o (vertical)
                             Focus range               8 m to ∞




view. Controlled experiments have shown that the driver often feels more comfortable looking at the night
vision display than at the actual road scene [3], which is a behavior that may have a negative influence
on the safety margin. Reliable automatic detection and moose early warning by aid of processing the
video signal is thus of interest.
  The similarities between the pedestrian and moose early warning systems are quite obvious; however,
there are some differences as well. The latter is a passive warning system with the major aim of detecting
big game hundreds of meters in front of the car.


D. Aim and goal of the paper

  The primarily goal of the paper is to present the equipment that has been developed, the moose
simulator, and the design of the field tests in order to create a public database useful for the research and
development of moose early warning driver assistance systems. Further, we report our practical findings
that have to be considered before applying a moose early warning system at large scale, as well as
give a review of relevant background information that a priori may be merged into such an advanced
driver assistance system, such as the spatial and temporal distributions of encounters with moose. A full
examination of this latter topic is, however, beyond the scope of this paper, in which we have restricted
the overview to the situation in Sweden.
  In our work, a commercial night driver system is patched to serve as the equipment basis for the
collection of a database for the design and evaluation of big game early warning systems. The design
and performance of a controllable living animal moose simulator is included as well. In Sec. II the
technical platform, the moose thermal replica, and the in-house developed moose early warning system

December 18, 2006                                                                                            DRAFT
                                                                                                                             6




Fig. 3.   The Nightdriver far infrared camera mounted in the front of the Audi S4 test vehicle. The size of the camera is L: 15
cm × W: 17 cm × H: 11 cm.




are described in some detail. Further, several typical traffic scenarios including a vehicle and a moose
are designed and form the basis for collection of the database during full scale field tests. In Sec. III, the
design of the field test as well as the database is described. The performance of an in-house developed
early warning system is illustrated in Sec. IV. Section V includes some practical considerations for these
kind of advanced driver assistance systems. The conclusions are drawn in Sec. VI.


                                                     II. E QUIPMENT

A. Night vision system

   The employed system is based on the Nightdriver system by L-3 Communications Infrared Products
— a third party product suited for the Hummer H2 [2]. For easy reference, the technical specifications
for the system are summarized in Table I. The system consists of the actual infrared camera (as shown
mounted on the test vehicle in Fig. 3) and a head unit equipped with the display for the driver, as well
as a factory outlet for the composite video signal. The Hummer H2 is not a typical private car, due to its
size and weight. In order to obtain video recordings typical for a private car, the employed test vehicle
was selected among typical mid-size private cars. The mounting of the camera unit in the front area of
the Audi S4 test car (B5-body, model year 2000) is shown in Fig. 3. The camera is aligned by hand
to set the view in the direction of the car. No particular calibration of its alignment is necessary. The
mounting of the camera is thus similar to the mounting employed in [5], where a correct localization
calibration procedure is presented.


December 18, 2006                                                                                                      DRAFT
                                                                                                                              7




Fig. 4.   Night vision equipment mounted inside the car compartment. From the top to bottom: head unit with head-up display,
extra display, digital camcorder with auxiliary video input, video splitter, and CB-radio for personal communication during field
tests.




   The camera video signal is transferred by wire to the head unit mounted inside in the car compartment
on top of the dashboard (see Fig. 4). The head unit provides visual output by a built-in head-up display,
but also by an NTSC video output signal. With reference to Figures 4–5, the video output from the
Nightdriver system is fed to an external display, a digital camcorder, and the video input on the PC is fed
through a video signal splitter. The Nightdriver system, video splitter, external display, and camcorder
are all supplied with power by the 12 volt system of the car. It is accentuated by the fact that all of the
utilized equipment consists of standard consumer electronics.
   An exemplary recording from the field test is shown in Fig. 6. A second digital camcorder is mounted
between the front seats and provides a reference recording of the view in front of the vehicle.


B. A moose thermal replica

   The moose is a wild animal and cannot easily be used for controlled experiments such as the one
presented in this work. A moose thermal replica may be designed and implemented by several means,
for example, by water-filled barrels mounted on a stand.


December 18, 2006                                                                                                       DRAFT
                                                                                                                            8

                                                      SIGNAL PROCESSING FLOW
                                                           RAW INFRARED DATA           PROCESSED DATA

                                                           HUD
                                    FIR
                                                      Head unit       Camcorder
                                    camera


                                                                        Display
                                                                                        Display

                                                                                  Laptop PC


                                                      REFERENCE VIDEO RECORDING
                                          Camcorder




Fig. 5.   Block diagram of the night vision driver assistance system utilized for the field tests. The system consists of the
Nightdriver camera and head unit with head-up display (HUD), camcorder, extra display, and laptop PC. A second camcorder
is used to record the reference video. In addition, GPS and maps are used for positioning and CB-radio for communication
between the vehicle, the moose simulator, and the field test director.




Fig. 6.   Original output of the employed night driver system showing the thermal signature of the moose simulator (Equina
Islandica) and its rider during practical field tests. The distance between the camera and horse is 92 meters measured by GPS.




   Here, a more pedestrian approach is employed and the moose is replaced by an Icelandic horse (that
is, Equina Islandica) and its rider. The weight of the Swedish moose typically spans the interval 200 to
550 kilograms, whereas the height from withers to front hoof may be up to 2 meters, where the bull is
20 % larger than the cow. The employed horse and carriage weighs some 340 kilograms and the height
from withers to front hoof is 1.38 meters. Thus, its size and weight correspond to a tiny to mid-size
moose. The moose, like the horse, has three gaits: walk, trot, and canter (although the Icelandic horse
also has two more gaits: tolt and pace). It mostly moves in walk but runs in trot if it has to cover longer
distances fast. The top speed of a moose in trot is 60 kilometers/hour, which it can keep up for up to 500

December 18, 2006                                                                                                     DRAFT
                                                                                                         9



meters at a time. The moose canters only on rare occasions, and in these cases, only for short distances.
For that reason, only walk and trot were considered for the field tests.
  Some advantages of the employed moose simulator over synthetic ones include its agility over difficult
terrain, similarity with the moose in cross-country running, and the natural temperature control of the
horse’s body. A field test thermal image of the simulator is shown in Fig. 6.
  All of the field tests were carried out in the natural environment of the horse under decent weather
conditions employing an experienced and skilled rider. It is stressed that during the experiments the
contracted horse did not suffer any discomfort, distress, pain, or injury. The horse was also accustomed
to traffic and it was not startled by the speedy movement or the roaring sound of the test vehicle during
the tests.


C. AME — moose early warning system software

  A moose early (AME) warning algorithm has been developed in-house which calculates a direction of
moose risk index based on the infrared video signal [17]. It is running in real-time with a low inherent
time delay, and is thus a suitable tool for a variety of tests spanning from algorithm development to
studies of the behavior of the driver. AME includes features like adaptive detection of moose-like objects
and Bayesian based dynamic tracking, and in addition it calculates a risk index value which invokes
a soft early warning signal to the driver, including an estimate of the relative direction of the moose.
Exemplary AME results are reported in Sec. IV; see [17] for a detailed description of functionality and
evaluation of its performance. In addition, the system may run in log mode to record all detection and
tracking activities for further development. By default, all recorded video is compressed with a standard
MS-MPEG4 V2 coder, and is resized to 640 × 450 pixels in order to remove the intrinsic blinking edges
of the video input. The AME software is also a versatile tool for collecting new experimental data in a
convenient manner.
  The functionality is implemented in C++ and the software is compiled for a standard personal computer
equipped with in- and output for analog video under Windows XP. The AME PC software is freely
available for non-commercial use at the AME site1 , hosted by the Royal Institute of Technology. The
AME software is being ported to a stand-alone Texas Instrument DSK module equipped with a video
daughter card, which provides a mechanically robust solution suitable for long-term field tests.

  1
      The AME software may be download from www.ame.ee.kth.se.




December 18, 2006                                                                                   DRAFT
                                                                                                         10


                                                   TABLE II

                         W EATHER CONDITIONS AT   THE FIELD TESTS DESCRIBED IN   F IG . 7.




                                    Time                      10 a.m.
                                    Condition                 mostly cloudy
                                    Temperature               11o C
                                    Chance of precipitation   70%
                                    Wind                      8 m/s
                                    Humidity                  94%




                            III. F IELD   TESTS AND   AME     PUBLIC DATABASE

A. Controlled test scenarios

  The recorded data originate from extensive use of the infrared imaging system at ordinary driving
situations, as well as from controlled field tests. The aim of the latter experiments is to specify the
scenarios for generality so that the most encountered situations are covered, for which a moose early
warning system may provide additional input to the driver, resulting in an increased safety margin.
Besides the recordings from the infrared camera, some reference video is recorded, GPS positioning and
estimation of moving patterns are performed, and recordings are made of weather conditions and other
pertinent information.
  In order to simplify the directed scenarios, additional constraints are imposed. Within the view angle
of the camera, no more than two heated bodies are present: the moose thermal replica and a pedestrian or
stationary object. Further, moving objects are constrained to a constant speed and straight line movements.
The field tests are carried out on a road at a horizontal open piece of ground without obstacles hiding
the visual view. The design of the field tests is based on observing some 20 hours of typical real-life
infrared recordings from Swedish country roads and highways.


B. Weather conditions

  The weather conditions are important for the recording of a database like the one presented here.
The quality of the infrared video signal depends not only on the temperature difference between the
warm objects and the surrounding, but also on surface properties of the heated body such as emissivity,
reflectivity, and transmissivity. According to the statistics presented in Fig. 2, the number of reported

December 18, 2006                                                                                   DRAFT
                                                                                                                              11




Fig. 7.     Schematic design of the nine scenarios for controlled experiments with the moose thermal replica and additional
heated-body obstacles. The rectangle indicates the test vehicle, the triangle indicates the moose thermal replica, and the circle
and the hexagon indicate a hot object and a pedestrian, respectively. All arrows represent constant speed.




accidents with moose obeys a visually periodic variation over the year. In general, most accidents are
reported during October, but in some years January also has a high number of reported accidents; mainly
depending on the snow conditions. The large number of reported accidents during the fall coincide both
with the mating season (that is, the season of the moose, not the car drivers) and with the increased
alertness of the moose due to the number of moose hunters and berry/fungus pickers in the woods during
the fall.
   Based on the presented statistics, and in order to correctly mimic the working-environment of the
AME system, the date of the field tests was chosen to be mid-October. The actual weather conditions
are displayed in Table II. According to Table II, the air temperature was low enough to result in a high
contrast between hot objects and the background through an infrared captured image. Test records were
only taken without rainfall (see the detailed discussion on the effect of rainfall in Sec. V).


C. The nine test cases

   The designed scenarios are summarized in Fig. 7. In total, nine scenarios were designed for the
controlled experiments, from a basic scenario with a stand-still vehicle and the moose thermal replica,
to a turning car with a running moose and walking pedestrian in and across the view of the driver.
The scenarios try to model the scenarios in which a moose early warning system should indicate some


December 18, 2006                                                                                                        DRAFT
                                                                                                           12



degree of danger. Simpler scenarios are employed as well in the application development phase, and
more complicated scenarios can be used for general testing purposes.
  Fig. 7 summarizes various designed scenarios. In scenario 1, the moose (thermal replica) is at a standstill
or perpendicularly crosses with a constant speed the view of the standstill FIR-camera. This essential
scenario is useful as a baseline to implement and verify different detection and tracking algorithms. In
scenarios 2 and 3, some disturbances, such as thermal obstacles and pedestrians, are introduced. These
scenarios are mainly used to test and develop morphology-based detection algorithms.
  In scenarios 4 and 5, the moose moves toward the vehicle, and in the latter scenario a moving pedestrian
is added. Compared to the previous scenarios, these scenarios result in gradually increasing the size of
the objects in view of the FIR-camera. By tracing the changes in the baseline of object bounding boxes
and the changes of their sizes, more detailed information, such as instantaneous speed of the objects and
relative distance between the objects and the test vehicle, can be estimated. This information acts as an
important parameter in evaluating the risk index, and also provides strong evidence in the morphology-
based detection algorithm.
  In scenarios 6 and 7, background motion is introduced by driving the test vehicle at a constant speed.
These tests are useful in the evaluation of motion-based detection algorithms. As they are more like a
real-life situation, these background motion scenarios are also useful resources in testing the robustness
in all cases. In scenarios 8 and 9, an extra but dominating background motion speed is introduced by
turning the test vehicle. These cases are useful to verify the robustness of the motion-based detection
algorithm. Similar to scenarios 6 and 7, these may also be used as test resources to all algorithms.


D. AME public database

  One goal of the presented project is to build up a database with recordings suitable for education,
research, and development around moose early warning systems. The database is publicly available
through the AME website2 .
  Besides the sequences from the nine controlled experiments reported above, public access is allowed
to a plurality of sequences from different driving scenarios, weather conditions, and time of day at a
variety of geographical locations.

  2
      The AME data-base is available at www.ame.ee.kth.se.




December 18, 2006                                                                                     DRAFT
                                                                                                                         13




Fig. 8. Detection algorithm results. Left column shows the original FIR captured images whereas right column shows the parts
of the target objects bounding box.




                    IV. I NITIAL      EXPERIMENTS WITH A MOOSE EARLY WARNING SYSTEM

  In this section, the performance of the in-house developed AME software is illustrated in some detail.
The underlying video processing is beyond the scope of this presentation and is reported in detail in [17].
The AME software, which includes features like threshold detection and continuously adaptive mean-shift
tracking algorithm, is available through the AME website 3 .
  Despite of using only pixel-wise information, threshold detection algorithm gives acceptable result in
identifying target objects. As shown in Fig. 8, targets or their composing elements are highlighted and
registered. The tracking algorithm evaluates the average relative speed of the target objects, which is an
important factor in risk evaluation. In order to calculate the risk index, parameters such as the relative
distance to target objects, position of the bounding box and the speed of target objects are used. As
shown in Fig. 9, the calculated risk index of each target object is represented by the size and position of
the corresponding round dot, respectively. When the risk index associated to a target exceeds a certain
predefined threshold, a frequent blinking of the round dot warns the driver. In order to present the
driver with a more comfortable and clearer view centric position of target objects are transformed to the
distance-angular dimension space under the assumption of a flat road in front of the vision system. The
AME software is still being developed with additional features and more robust functions. In an upcoming
release, a morphological module or its equivalence will be added to refine the threshold detection criterion,
which makes the total algorithm more reliable and delivering less false alarms [17].

  3
      The AME software is available at www.ame.ee.kth.se



December 18, 2006                                                                                                    DRAFT
                                                                                                                       14




Fig. 9. Exemplary human-vehicle interface of the AME moose early warning system. The round dots represent potential risks,
where the location and size indicate the position and size of the heated body targets.




                                          V. P RACTICAL       CONSIDERATIONS

  In this Section, some practical considerations are discussed that have to be considered before applying
a moose early warning system in large scale or integrating the warning system with existing technologies
like GPS and electronic stability programs. The discussion is based on an extensive use of the descried
technical equipment during a 6 month period of time. The practical use includes several encounters with
wild moose, as well as other wild game, domestic cattle, pedestrians, et cetera. The tests include locations
such as highways as well as public roads at different times of day and for a large variety of weather
conditions.


A. Effects of rain

  A military countermeasure to hide the thermal signatures of heated bodies is the use of water drops,
more specifically so-called multi-spectral water fog produced by a mixture of water drops of different sizes
[16]. Clearly, a similar but weaker effect is obtained when using night vision-type equipment during rainy
weather. In such scenarios, the video signal obeys inferior resolution and contrast compared with mainly
dry weather conditions. Accordingly, the detection probability of the early warning will significantly drop,
basically down to zero at heavy rain.


B. Camera dirtying

  The front mounted infrared camera is vulnerable against dirtying that will significantly reduce the
performance of the system over time. In particular, rainy weather typically implies an increased camera

December 18, 2006                                                                                                  DRAFT
                                                                                                          15



dirtying. Practical experiments show that dirtying has to be handled by, for example, a water-based
cleaning system similar to the one compulsory in some countries for cleaning the headlamps.

C. Warm rock image clutter

  Our long-term practical countryside use of the night vision system has pointed out one significant risk
of false alarms. Stones and rocks that have been heated during warm days with high solar radiation show
a significant heat signature for a long period of time. When twilight sets in, the warm stones and rocks
result in the most pronounced heat signatures. Accordingly, an effective early warning system must be
able to handle not only warm objects, but also provide measures in order to reduce the false alarm rate
due to stationary warm objects. For urban environment, the amount of image clutter increases due to
vehicles, transformers, electric boxes, et cetera [4].

D. Effect of a narrow beam width

  The employed system has a narrow beam-width, and thus it is basically able to provide early warnings
for moose in the far-field in front of the vehicle. The implication is that time for extra signal processing
may be utilized in order to reduce the risk of false alarm. The disadvantage is that the system is not able
to react on moose that cross the road in the immediate front of the vehicle — a crossing that with high
probability will end in an encounter with severe injuries. Thus, a moose early warning system relying
on vehicle-based infrared imaging is not the one and only solution for fulfilling a zero-vision in number
of deaths.

E. Man-vehicle interface

  Controlled experiments show that drivers may spend a considerable amount of time looking at the
night vision display [3]. In [3], figures up to 75% of the time were reported for some test persons.
Our long-term use of the system verifies this finding; from a driver point of view one often feels more
comfortable looking at the head-up display than at the actual road scene. As pointed out in [3], such a
behavior may have a negative influence on the safety margin. The countermeasure proposed was simply
to make the display unlit during uneventful periods of time, thus forcing the attention of the driver to
the actual road scene.
  Our observations also indicate that drivers lose interest in viewing the display as time goes by, and thus
critical information may be overlooked. It is also observed that tired drivers may get confused shifting
their view between the road scene, rear-view mirrors, and infrared display. Thus, automatic detection by
aid of signal processing is a powerful tool for alerting the driver and keeping his attention on the road.

December 18, 2006                                                                                    DRAFT
                                                                                                        16



F. Exemplary data fusion for improved estimation of relative risk

  In our approach, the risk index estimator is solely based on the information extracted from the infrared
camera video signal. This is a crude estimation of the relative risk, which may be improved by fusion
information from additional sensors and pertinent prior knowledge. One such additional sensor is a
GPS-receiver (or another satellite-based system) for positioning of the vehicle in time and space. Since
the temporal as well as the spatial distribution of the fatal moose - personal car encounters are well-
reported (at least in Sweden, in the STORM database the monthly encounters are resolved into 21
administrative provinces), it is straightforward to weight the information extracted from the video signal
with the geographical accident density at a given time in order to get an improved risk estimate. It
is straightforward to include statistics for other wildlife reported in STORM. Besides bears, wolves,
wolverines, lynxes, moose, deer, roe deer, otters, wild boars, mouflons, and eagles, STORM reports
encounters with reindeer, which are categorized as domestic cattle. In particular, this fusion of data may
reduce the risk of false alarms — however, one should realize that on occasion the moose may be spotted
in the most unexpected places.


G. Integration with electronic stability programs

  Traffic safety is tested by institutes and organizations, but also more informally for the public, for
example by motor magazines and newspapers. The Swedish ‘moose-test’ (in German ’Elchtest’) is a test
to check the ability to swerve and avoid a collision. It became well-known after the failures in test by
a new model on the market. The purpose of the test is to simulate an encounter with the moose, with
a typical evasive maneuver — a maneuver all cars should pass at reasonable speeds and specifically
without overturn. Electronic stability programs (ESPs) are a key tool for increased traffic safety where
the stability of the vehicle is controlled by feedback of sensor signals from gyroscopes, accelerometers,
et cetera in order to control the brakes on the individual wheels. Clearly, a moose early warning system
can provide feed forward information to such a stability program for advanced control strategies when
an evasive maneuver can be predicted.


                                          VI. C ONCLUSIONS

  In this paper a framework for moose early warning driver assistance systems has been presented,
including publicly available baseline AME software and a database suitable for education, research, and
development within the area. The cost for the infrared imaging equipment is high compared with the



December 18, 2006                                                                                   DRAFT
                                                                                                                            17



additional cost related to the implementation of an automatic moose early warning system, an add-on
feature that may increase the safety margin ever further.


                                                 ACKNOWLEDGMENTS

  The authors wish to acknowledge Kjell Norén for design and manufacturing of the facilities needed to
mount the Nightdriver system at the employed test vehicle, Tomas Andersson for providing photographs,
Andreas Stenhall for maintaining the AME web-site; all at the Royal Institute of Technology. Further, Lars
Sävberger at the SES-gruppen for providing data from the STORM database, Älgskadefondföreningen
for providing background material, Jonas Nordenmark at Lava Electronics AB for technical support
and providing hardware, Glödlampsfabrikernas stipendiefond for partial funding of the hardware, Sofies
Islandshästar for lending us horse and facilities, and Dr Daniel Rönnow for sharing his expertise on
infrared imaging. Finally, Magni frá Stóra-Sandfelli is acknowledged for being an excellent moose
simulator.


                                                       R EFERENCES

 [1] Road Safety and Environmental Benefit-Cost and Cost-Effectiveness Analysis for Use in Decision-Making – Examples of
     Assessed Road Safety Measures, a Short Handbook, European Commission, July 2006.
 [2] Raytheon Owners Manual Nightdriver for Hummer H2, model # 3243777, Raytheon Commercial Infrared, 2003.
 [3] R. Kovordanyi, T. Alm and K. Ohlsson “Night-vision display unlit during uneventful periods may improve traffic safety,”
     Intelligent Vehicles Symposium 2006, June 13-15, 2006, Tokyo, Japan, pp. 282-287.
 [4] Y. Fang, K. Yamada, Y. Ninomiya, B.K.P. Horn and I. Masaki, “A shape-independent method for pedestrian detection with
     far-infrared images,” IEEE Transactions on Vehicular Technology, 53(6):1679-1697, Nov. 2004.
 [5] M. Bertozzi, A. Broggi, A. Fascioli, T. Graf and M.-M. Meinecke, “Pedestrian detection for driver assistance using
     multiresolution infrared vision,” IEEE Transactions on Vehicular Technology, 53(6):1666-1678, Nov. 2004.
 [6] F. Xu, X. Liu and K. Fujimura, “Pedestrian detection and tracking with night vision,” IEEE Transactions on Intelligent
     Transportation Systems, 6(1):63–71, March 2005.
 [7] X. Liu and K. Fujimura, “Pedestrian detection using stereo night vision,” IEEE Transactions on Vehicular Technology,
     53(6):1657-1665, Nov. 2004.
 [8] M. Bertozzi, A. Broggi, A. Lasagni, and M. Del Rose, “Infrared stereo vision-based pedestrian detection,” IEEE Intelligent
     Vehicles Symposium, Las Vegas, Nevada, June 6-8, 2005, pp. 24-29.
 [9] F. Suard, A. Rakotomamonjy, A. Bensrhair and A. Broggi, “Pedestrian detection using infrared images and histograms of
     oriented gradients,” Intelligent Vehicles Symposium 2006, June 13-15, 2006, Tokyo, Japan, pp. 206-212.
[10] T. Tsuji, H. Hattori, M. Watanabe and N. Nagaoka, “Development of night-vision system,” IEEE Transactions on Intelligent
     Transportation Systems, 3(3):203–209, Sept 2002.
[11] A. Shashua, Y. Gdalyahu and G. Hayun, “Pedestrian detection for driving assistance systems: single-frame classification
     and system level performance,” IEEE Intelligent Vehicles Symposium, Parma, Italy, June 14-17, 2004, pp. 1-6.



December 18, 2006                                                                                                      DRAFT
                                                                                                                            18



[12] B. Fardi, U. Schuenert and G. Wanielik, “Shape and motion-based pedestrian detection in infrared images: a multi sensor
     approach,” IEEE Intelligent Vehicles Symposium, Las Vegas, Nevada, June 6-8, 2005, pp. 18-23.
[13] M. Bertozzi, A. Broggi, M. Felisa, G. Vezzoni and M. Del Rose, “Low-level pedestrian detection by means of visible and
     far infrared tetra-vision,” IEEE Intelligent Vehicles Symposium, June 13-15, 2006, Tokyo, Japan, pp. 231-236.
[14] A. Broggi, R.I. Fedriga, A. Tagliati, T. Graf and M. Meinecke, “Pedestrian detection on a moving vehicle: an investigation
     about near infrared images,” IEEE Intelligent Vehicles Symposium, June 13-15, 2006, Tokyo, Japan, pp. 431 - 436.
[15] SYSCO Today, January 2005 Canadian Edition, p. 28.
[16] G. Olsson, “Multispectral waterfog as countermeasure,” Nordic Symposium on Military Electro-optics, October 17-18,
     2001, Helsinki, Finland.
[17] Y. Yao, “Moose early warning driver assistance system," Technical report, Royal Institute of Technology, Stockholm,
     Sweden.




December 18, 2006                                                                                                      DRAFT
                                                                                                             19



                                            L IST   OF   F IGURES

1      Typical outcome of an encounter with a Swedish moose. The roof at the A-frame has been
       exposed by a significant down-force by the trunk of the moose. The front of the private
       car is rather unaffected by the encounter. The moose died. Photo: S. Sundkvist, courtesy of
       Älgskadefondföreningen. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .     2
2      Statistics from the police authority STORM database showing the number of reported car
       accidents with moose in Sweden for the time period of January 2003 to October 2006. The
       yearly peaks in October are marked with ’o’. . . . . . . . . . . . . . . . . . . . . . . . . . .      3
3      The Nightdriver far infrared camera mounted in the front of the Audi S4 test vehicle. The
       size of the camera is L: 15 cm × W: 17 cm × H: 11 cm.           . . . . . . . . . . . . . . . . . .   6
4      Night vision equipment mounted inside the car compartment. From the top to bottom: head
       unit with head-up display, extra display, digital camcorder with auxiliary video input, video
       splitter, and CB-radio for personal communication during field tests. . . . . . . . . . . . . .        7
5      Block diagram of the night vision driver assistance system utilized for the field tests. The
       system consists of the Nightdriver camera and head unit with head-up display (HUD),
       camcorder, extra display, and laptop PC. A second camcorder is used to record the reference
       video. In addition, GPS and maps are used for positioning and CB-radio for communication
       between the vehicle, the moose simulator, and the field test director. . . . . . . . . . . . . .       8
6      Original output of the employed night driver system showing the thermal signature of the
       moose simulator (Equina Islandica) and its rider during practical field tests. The distance
       between the camera and horse is 92 meters measured by GPS. . . . . . . . . . . . . . . . .            8
7      Schematic design of the nine scenarios for controlled experiments with the moose thermal
       replica and additional heated-body obstacles. The rectangle indicates the test vehicle, the
       triangle indicates the moose thermal replica, and the circle and the hexagon indicate a hot
       object and a pedestrian, respectively. All arrows represent constant speed. . . . . . . . . . . 11
8      Detection algorithm results. Left column shows the original FIR captured images whereas
       right column shows the parts of the target objects bounding box. . . . . . . . . . . . . . . . 13
9      Exemplary human-vehicle interface of the AME moose early warning system. The round
       dots represent potential risks, where the location and size indicate the position and size of
       the heated body targets.   . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14




December 18, 2006                                                                                       DRAFT

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:29
posted:2/23/2010
language:English
pages:20