Post on 10-May-2020
transcript
ANALYSIS OF LIDAR SENSORS FOR NEW ADAS
APPLICATIONS. USABILITY IN MOVING OBSTACLES
DETECTION.
Authors: F. García
1*, F. Jiménez
2, J.E. Naranjo
3, J.G. Zato
3, F. Aparício
2, J.M.
Armingol1, A. de la Escalera
1.
1 Universidad Carlos III de Madrid. Laboratorio de Sistemas Inteligentes
Avda. De La Universidad 30, 28911 Leganés (Madrid). Spain. Telephone: +34 91 624 99 70. E-mail: fegarcia@ing.uc3m.es
2 Universidad Politécnica de Madrid. E.T.S.I. Industriales. INSIA
Carretera de Valencia, km.7, 28031 Madrid. Spain. 3 Universidad Politécnica de Madrid. E.U. de Informática
Carretera de Valencia, km.7, 28031 Madrid. Spain.
ABSTRACT
ADAS applications feasibility is connected with the necessity of trustable sensors. The lack
of cheap and reliable sensors underlines the need to use different sensors at the same time.
LIDAR provides reliable but limited information of the surroundings for a vehicle
application. This paper presents a comparison between two different kinds of LIDAR
sensors focusing on their possibilities of being used in ADAS applications. Finally a new
method for detecting moving obstacles, mainly vehicles, is proposed. This method has been
implemented and tested; results of the different test performed are shown.
KEYWORDS
Data Fusion, Intelligent vehicles, ADAS, LIDAR
1.- INTRODUCTION
ADAS applications feasibility is connected with the necessity of trustable sensors. In this
context, the lack of cheap and reliable sensors underlines the need to use different sensors
at the same time in order to provide a reliable and accurate application. A possible set of
sensors used in data fusion applications are computer vision and LIDAR. The reason for
using these sensors is that LIDAR provides a reliable source of possible detections in the
surroundings; on the other hand, data provided by vision sensors allow the different objects
detected by the LIDAR to be classified. This paper focuses on the necessity of a reliable
LIDAR sensor, able to detect the surroundings obstacles and even been able to give a first
estimation of the shape of the detected obstacle.
Two different LIDAR have been tested in different conditions and movements to provide a
comparison of their possibilities for ADAS applications. Typically, LIDAR are used to
provide detection, but not usually give an estimation of the kind of obstacle that is detected.
A novel method obstacle detection and classification is proposed, this classification focus
on the most important obstacles that can be found in a road environment, pedestrian and
vehicles. These obstacles represent other users of the road, which detection and tracking are
crucial for developing safety applications. This information can be lately used in a fusion
application together with visual sensors to perform a more accurate classification.
Radar researches are widely used in applications for road environments. In [1] frequency
laser radar are used to classify vehicles according to their height, [2], [3], [4] and [5]
combines frequency radar applications and other sensors like visual information.
In the latest years, LIDARs are becoming more popular in road environment applications,
[6] performs classifications and tracking based in different possibilities for each detected
obstacle. [7] and [8] detects and classificates pedestrians based on the movement of the
obstacle. [9] detects potentially dangerous zones in the road where pedestrians are more
vulnerable by computing several sequences and with obstacle detection correlation. [10]
and [11] Detects and classifies based on the shape and movement.
Other ADAS application which does not involve obstacle classification can be developed
using laser range sensors [12] and [13].
2.- OBJETIVES AND METODOLOGY
The following objectives have been considered:
- Compare the detection using two different sensors and establish the detection and
identification limits for each sensor studied. The following equipment was tested LD
LRS 1000 and LMS-291 both of them provided by SICK. Both of them have different
qualities and belong to two different LIDAR families.
These tests were performed in a controlled environment which could be configured
specifically for these tests. And the vehicles that were used were a metallic-grey
Peugeot 307 and a black Nissan Note (Figure 1). The first vehicle represents the best
case scenario were the reflectivity is high, thus no losses due lack of reflectivity can be
found. The second vehicle is the worst case scenario; its black paint represents the most
challenging situation for the laser range finder due its lower reflectivity. So in this test,
not only a comparison between two families of laser range finder LIDARs are
performed, also a the viability of the laser radars in ADAS applications are tested in the
worst case scenarios.
Figure 1. Test vehicle with radars mounted in the bumper (center), black colored car
(left) and gray colored car (right).
- A novel approach to vehicle and pedestrian detection in road environments presented in
[14] and tested in different conditions, results are shown.
3.- LIDAR COMPARISION
The sensors used are both 2D LIDAR from SICK. They were mounted in the bumper of a
vehicle to perform several test sequences in order to compare the capabilities of each sensor
(Figure 2). Before the main part of the test is explained, the main characteristics of each
sensor are presented.
LRS-1000
The data acquisition frequency of the LRS-1000 can be selected from 5 to 10 Hz. Its
maximum distance measured is up to 250 meters and the resolution can be selected from
0.125 to 1.5º. The LSR-1000 is a high profile laser measurement system, its wide field of
vision, up to 360º, its high maximum distance and lower resolution makes it a very
interesting tool for ADAS application. The main disadvantage that can be found in this part
is the lower frequency; it gives a scan every 100 msecs.
LMS-291
Can be configured with a frequency up to 76 Hz. but for the selected resolution (0.25º) its
maximum operating frequency is set to 19Hz The resolution can be selected from 0.25 to
1º. And its maximum detection distance is 80m. The main characteristic is its high
frequency, it gives information every 52msecs. In ADAS applications, real time detection is
a critical point, thus the faster the detection the more suitable the sensor for these
applications.
Figure 2. LMS 291 (left) and LSR 1000 Laser Radar (rigth).
Table 1. A priori comparison of laser performance.
LRS-1000 LMS-291
Field of Vision 360 180º/100º
Resolution 0.125º to 1.5º 0.25º to 1º
Max Distance measure 250m 80m
Detection Distance(1)
229,2 m >80 m
Max Distance(2)
114.6 m 57.3 m
Working frequency 10Hz 19Hz
(1) At least four detection point, from the formula:
22 mintg
ddist , for a car with 2 meters width means
d=0.5m (distance between points). (2)
It has been estimated that 8 points is the minimum amount of points necessary to give a good estimation.
For a 2 meters width car, d=0.25.
In order to evaluate the accuracy of the measurements, different Tests have been
performed. Tested movements are shown in figure 3.
Figure 3. Test performed. Separating movement A, Approaching movement B and
crossing movement C.
In each movement, detection widths for the moving vehicle were recorded, as well as the
number of detected points that defines the car. Results are shown in figure 4 to 8:
Figure 4. Width measure for LRS-1000 movements A and B in millimeters.
Figure 5. Width measure for LMS 291
movements A and B in millimeters.
Figure 6. Pulses detected for LRS-1000
movements A and B.
0
20
40
60
80
5 15 25 40 60 80 100 120
Distance detected [m]
Width[m]
in
meter
s
Width[m]
in
meter
s
Distance detected [m]
in meters
# of points detected
in
meters
Distance detected [m]
in meters
Figure 7. Pulses detected for LMS 291
movements C.
Figure 8. Pulses detected for LMS-291
movements A and B.
Real widths measures given by manufacturers are:
Nissan Note= 1691 millimeters.
Peugeot 307= 1762 millimeters.
Test conclusions:
As it is shown, width accuracy is very similar for both sensors, but LMS 291 gives better
results at long distances. And the number of points detected for the desired configuration
gives also very similar configuration. Although laser radar LRS 1000 gives results in
distances higher than 80 meters.
LRS-1000 gives higher capabilities which make it a very interesting sensor for long range
detections, due to its wide vision range and its extremely high resolution. Its lower
frequency and long distance measurements makes that more structured environments, with
less changing conditions, and where long distances are important, are typically the best
scenarios for this kind of LIDAR. Typically interurban scenario matches with these
requirements. Where detections has to be done at longs distances ( > 200 meters ) due to
the speeds of the cars involved. Closer cars, on the other hand, usually do not perform
special trajectories to be tracked, so real time tracking is not so important, but long distance
detections are.
The experiment performed with the LMS-291-S05 gives similar results, but in a lower
distance range. What makes interesting this model, in comparison to the LRS-1000, besides
its lower price, is its higher detection frequency. This lower frequency makes it a very
interesting solution, when dealing with extremely changing environments, as urban
scenarios, where usually vehicles have to deal with lower distances and fast changing
conditions. In urban scenarios, cars, bikes, pedestrian or other kind of obstacles can appear
from any direction with variable trajectories, so real time tracking is mandatory, to detect
dangerous situations and warn the driver. Thus in these scenarios a fast response is crucial,
so this sensors is the best solution in urban scenarios.
0
20
40
60
0 5 10 15
Black vehicle C movement for LRS 1000
Gray vehicle C movement for LRS 1000
Black vehicle C movement for LMS 291
Gray vehicle C movement for LMS 291
0
20
40
60
5 15 25 40 60 80
Black vehicle B movement
Black vehicle A movement
Gray vehicle B movement
Gray vehicle A movement
# of points detected
in
meters
Distance detected [m]
in meters
Distance detected [m]
in meters
# of points detected
in
meters
It also has been proved that detection can be performed even in worst case scenarios where
detected cars are black. According to figures 6, 7 and 8, points are enough to detect with
reliability even most difficult obstacles. Results also shown that it is easier detect moving
obstacles when they are performing separating movements that approaching movements.
The reason for such a special behavior is that front of the car presents a special
configuration that makes more difficult the reflex ion that la rear part. Radiator and other
parts of the front of the car present special reflexions that makes harder laser detection.
4.- MOVING VEHICLE DETECTION ALGORITHM
LIDAR main disadvantage is the relative low information provided, but enough to give a
first estimation of the shape of the obstacle detected and even provide moving vehicle and
pedestrian detection.
An application has been developed in the scope of this test to provide low obstacle
detection and identification [14]. The algorithm consists on a low level detection and
identification, and higher level tracking. This tracking stage not only records and predicts
the movement of the vehicles and pedestrian, also is useful to give a more accurate
detection.
The detection points given by the laser are merged together and labeled, representing
different obstacles (Figure 9). Polylines are created to join the different points that represent
the obstacles [9]. After some line merging, each obstacle is represented by a single polyline
proportional to its shape. 5 sets of obstacles were defined (L Shaped, Fixed Obstacle,
Road Borders, Possible Pedestrian and Moving Obstacle); each of them represents
different objects that can be differentiated in road environment using the information
provided by laser radar.
Figure 9. Different Obstacle Classification.
Moving obstacles
The pattern given by these kinds of obstacles makes possible detect and track their
movement. This obstacle can be detected using LIDAR LMS 291 and its special behavior.
For 0.25 º detection, it performs 4 scans independently which give 4 sets of spots with 1º of
resolution. Each scan is separated 0.25º in relation with the previous one. So after 4 scans,
the LIDAR returns a complete set of spots separated 0.25º. When a moving obstacle is
found, the four scans performed by the LIDAR for a single detection appear with a
variation which is proportional to the speed and direction of the detected object and the test
vehicle. Measuring the distance between two consecutive points we can calculate the speed
of the car in m/s (Figure 10).
where T is the rotation period which is T=13msecs. As the number of scans is 4, three
speeds can be measured in order to provide a more reliable measure.
3
12
t
yy
Vy
nnN
Nn
, 3
12
t
xx
Vx
nnN
Nn
,
, where t=13 msegs and v is in
m/s.
Figure 10. Moving vehicle Pattern.
False positives can be avoided detecting impossible speeds or movements. LRS
information cannot give this pattern in a single spot detection, but by combining more than
one scan a similar pattern can be detected.
Tracking stage, computes the speed of the moving vehicle, first using the low level
calculated speed and lately using the high level velocity calculated using this higher level
tracking. The algorithm calculated the position of the car in the subsequent detection scan
and search within a given window for another obstacle whether moving or any other
obstacle. Moving obstacle movements are recorded. After some scans it is checked if the
car is actually moving or not and higher level classification is performed according to the
movement of the car. The higher level classification algorithm is based in a voting scheme
that uses the ten latest movements and low level classification to perform the final decision:
,
where Vi represents the number of votes for each kind of obstacle, and is the gain factor
for each obstacle, which is only different for road borders and moving obstacles.
Finally there is a correction factor that corrects the low level classification, if it was not
considered moving obstacle in low level detection and the detection window finds it inside
a moving window, the low level detection is corrected only if real movement is detected in
the obstacle within the last two sequences.
Window width= Window height =
Where K1, K2, Th1 and Th2 are configurable.
Test Performed:
Several experiments were performed to test the proposed method, a GPS sensor ASTECH
G12 GPS and a speed sensor CORREVIT L-CE were used to increase the accuracy of the
test.Some result of the test performed are shown in Figure 11. It shows the percentage of
moving obstacles to be detected according to the distance. Two movements were tested, a
vehicle moving in direction to the LIDAR (approaching movement), and in the opposite
direction (separating movement):
Figure 11 Test Performed to check the algorithm.
Experimental results are shown in figure 12 and 13:
Figure 12 Classification Results
Percentage of detection vs distance in
meters.
Figure 13: Detection Percentage. vs.
Distance in meters. Overall Results.
The probability of being detected is different if the car is separating or approaching. The
best probability of being detected is when the car is separating; mainly because the back of
the car is bigger than the front part of the car, so the detected part of the car has more
surfaces to be detected by the sensor avoiding errors due pitching movements.
The results presented in Figure 13 only focuses in the probability of moving obstacles to be
detected in a single scan. A subsequent integration along time with a tracking algorithm has
lead to a much better detection, been able to track vehicles with no misdetections.
Results given by Figure 14 shows that a car is detectable within 30/40 meters in
approaching movements and until 80 meters in separating movements, this lead to a reliable
algorithm which is able to track the movement of cars in short distance. Given the
frequency of the LIDAR sensor used (LMS 291) and detection ratio given in this test, the
application presented here proved to be very accurate. It is especially in urban
environments, where short distances and fast movements are common and speeds lower so
a car detection within 40/30 meters can help to warn drivers in case of hazardous situations.
Figure 14 Tracking, distance in meters since the moving car is detected.
5.- CONCLUSION AND FUTURE WORKS
The first conclusion that can be obtained is that LIDAR LMS-291 provides less detection
performance. Enough to give a good estimation of the surroundings of road environments,
0
50
100
82 67 52 37 22 7
B
A
D
C
F 0
50
100
80 68 56 44 32 20 8
separating movement
approaching movement
0
20
40
60
80
100
B A D C F G E H
% of detections
in
meters
Distance [m]
Distance [m] % of
detections
in
meters
Distance where tracking starts [m]
in meters
Movement
Approaching
movement
Separating
movement
but ineffective in long range detections, since it is not able to detect object over 80m in
distance. However its low cost and high frequency makes it a very interesting option for
low cost ADAS application.
LIDAR LRS-1000 is a good sensor for road environments mainly in interurban areas where
speeds are high and detections should be done in long range distances. In urban areas both
sensors are able to give information enough to give shape estimation and movement
detection, but acquisition frequency may be taken into account.
The algorithm is based in the operation principle of the LMS-291. But it can be used with
the LRS-1000 by integrating several scans and look for the pattern variation as it is done
with the LMS-291.
6.- ACKNOWLEDMENTS
The work reported in this paper has been partly funded by the Spanish Ministry of Science
and Innovation (SIAC project TRA2007-67786-C02-01, TRA2007-67374-C02-01 and
TRA2007-67786-C02-02) and the CAM project SEGVAUTO.
7.- REFERENCES
[1]. Ildar Urazghildiiev, Rolf Ragnarsson, Pierre Ridderström, Anders Rydberg, Eric Öjefors, Kjell
Wallin, Per Enochsson, Magnus Ericson, and Göran Löfqvist. “Vehicle Classification Based on
the Radar Measurement of Height Profiles”. IEEE Transactions on Intelligent Transportation
Systems, Vol. 8, No. 2, June 2007.
[2]. Hofmann, U., Rieder, A., Dickmanns, E.D.: Radar and vision data fusion for hybrid adaptive
cruise control on highways. In: ICVS '01: Proceedings of the Second International Workshop
on Computer Vision Systems, London, UK, Springer-Verlag, 125-138, 2001.
[3]. Ofer, A.S., Mano, O., Stein, G.P., Kumon, H., Tamatsu, Y., Shashua, A.: Solid or not solid:
Vision for radar target validation. In: IEEE Intelligent Vehicles Symposium Procedings, 819-
824, 2004.
[4]. Steux, B., Laurgeau, C., Salesse, L., Wautier, D.: Fade: a vehicle detection and tracking system
featuring monocular color vision and radar data fusion. Volume 2, 632-639, June 2002.
[5]. Alessandretti, G., Broggi, A., Cerri, P.: Vehicle and guard rail detection using radar and vision
data fusion. Intelligent Transportation Systems, IEEE Transactions on 8(1), 95-105, March
2007.
[6]. Daniel Streller, Klaus Dietmayer, Jan Sparbert. “Object tracking in traffic scenes with multi-
hypothesis approach using laser range images” Proceedings of ITS 2001, 8th World Congress
on Intelligent Transport Systems, Sidney, October 2001.
[7]. Kay Ch. Fuerstenberg, Ulrich Lages. “Pedestrian Detection and Classification by
Laserscanners”. In Procs. IEEE Intelligent Vehicles Symposium 2002.
[8]. Fuerstenberg, K. Ch.; Dietmayer, K. C. J.; Willhoeft, V.: Pedestrian Recognition in Urban
Traffic using a vehicle based Multilayer Laserscanner. Proceedings of IV 2002, IEEE
Intelligent Vehicles Symposium, IV 2002 Versailles, Paper IV-80.
[9]. Broggi, A., Cerri, P., Ghidoni, S., Grisleri, P., Jung, H.G.: Localization and Analysis of Critical
Areas in Urban Scenarios. In: Procs. IEEE Intelligent Vehicles Sympo- sium 2008, Eindhoven,
Netherlands (June 2008) 1074–1079.
[10]. Streller, Kay Furstenberg, Klaus Dietmayer. “Vehicle and object models for robust tracking in
traffic scenes using laser range images”. The IEEE 5th International Conference on Intelligent
Transportation Systems. 3-6 September 2002, Singapore.
[11]. Kay Ch. Fuerstenberg, Klaus C. J. Dietmayer, Stephan Eisenlauer, Volker Willhoeft.
“Multilayer Laserscanner for robust Object Tracking and Classification in Urban Traffic
Scenes” Proceedings of ITS 2002, 9th World Congress on Intelligent Transport Systems,
Chicago 2002.
[12]. Jan Sparbert, Klaus Dietmayer, Daniel Streller. “Lane Detection and Street Type Classification
using Laser Range Images”. 2001 IEEE Intelligent Transportation Systems Conference
Proceedings - Oakland (CA), USA - August 25-29, 2001.
[13]. Fuerstenberg, K. Ch.; Hipp, J.; Liebram, A. (2000) A Laserscanner for detailed traffic data
collection and traffic control. Proceedings of ITS 2000, 7th World Congress on Intelligent
Transport Systems, Turin, Paper 2335, 2000.
[14]. García, F., Cerri, P., Broggi, A., Amingol, J.M., de la Escalera, A.: Vehicle Detection Based
on Laser Radar. Eurocast 2009.