Sie sind auf Seite 1von 6

Proceedings of the 2006 IEEE/RSJ

International Conference on Intelligent Robots and Systems


October 9 - 15, 2006, Beijing, China

Learning Trafc Patterns at Intersections by


Spectral Clustering of Motion Trajectories
Stefan Atev, Osama Masoud, and Nikos Papanikolopoulos
Department of Computer Science and Engineering
University of Minnesota Minneapolis, MN

Abstract We address the problem of automatically learning measures with certain invariance properties. For example, [7]
the layout of a trafc intersection from trajectories of vehicles proposes a similarity measure for moving object trajectories
obtained by a vision tracking system. We present a similarity based on an extension of the symbolic Edit Distance to real se-
measure which is suitable for use with spectral clustering in
problems that emphasize spatial distinctions between vehicle quences. Related work based on the Discrete Time-Wrapping
trajectories. The robustness of the method to small perturbations approach is presented by [8], while methods based on the
and its sensitivity to the choice of parameters are evaluated using Longest Common Subsequence approach are discussed in [9].
real-world data. While the time component of vehicle trajectories is somewhat
relevant to our problem, it is of secondary importance, hence
I. I NTRODUCTION
our departure from time-series similarity measures.
Automatic identication and tracking of vehicles in clut- The spectral clustering method used in our work is derived
tered outdoor scenes is a challenging problem for which from the method proposed by [10], taking into account de-
different solution approaches have been proposed. Systems velopments by [11] who proposed the use of local scaling
such as those by [1][4] collect vehicle trajectory data, either to handle both multi-scale data and background clutter. The
as a by-product of their operation, or as their ultimate goal. asymmetric nature of the local scaling method is a natural
The focus of this work is on learning the layout and xed t for the directed Hausdorff distance, which we generalize
ingress/egress structure of a trafc intersection automatically, for use with vehicle trajectories. A similar modied Hausdorff
once moving object trajectories are extracted from a video clip distance allowing for partial matching of image edge templates
by any convenient method. was introduced by [12]. The partial matching formulation was
Knowledge of the xed intersection structure enables the developed further in [13] to increase the robustness of object
computation of statistics that are of interest to trafc engineers, matching in the presence of noise; the authors also investigate
such as: turning gap times for left and right turns, headway in detail various modications of the Hausdorff distance that
times, lane change and U-turn event counts, as well as detailed are related to our approach.
trafc volume information. Automatic identication of the The rest of this paper is organized as follows: In Section II
structure of an intersection reduces operator input required for we summarize our method for collecting trajectory data from
the use of trafc data collection systems and eliminates the video sequences. Section III presents our trajectory similarity
nuisance of manually modeling the structure of trafc scenes measure and discusses the motivation for its choice. Trajec-
of interest. While camera calibration is still required for our tory preprocessing and clustering using a spectral method is
trajectory collection method, the amount of manually specied described in Section IV. Experiments and results are discussed
data required for the operation of the data collection system in Section V, and our conclusions and future investigative
is signicantly reduced. endeavors appear in Section VI.
Our work bears most similarity to the approach of [5], where
spectral clustering is applied to extract trajectory patterns II. T RAJECTORY DATA C OLLECTION
using a similarity measure between Hidden Markov Model The vehicle tracking algorithm we employ is not the focus
parameterizations of the trajectories. Unlike [5], we measure of this work, but we will briey explain it for completeness of
trajectory similarity directly, not in a derived feature space; exposition. We collect trajectories using a region-based vehicle
further differences are in our use of automatic local scaling tracker capable of handling temporary static and dynamic
of the afnity matrix and in the determination of the optimal occlusions. All output variables of the tracker are reported
number of clusters. A paper by [6] investigates agglomerative in real-world units, using camera calibration data obtained
clustering of trajectories obtained by a multiple-hypothesis, by a primitive-based calibration procedure. In addition to
EKF-based tracker; the similarity measure is computed from position and velocity data, the tracker estimates the real-world
projections of trajectories on coordinate axes, and is closely dimensions of tracked objects, which enables the system to
related to methods used by the data mining community. differentiate between pedestrians and vehicles.
Methods for comparing multidimensional time sequences The tracking module uses an adaptive background model
are actively being developed for data mining and retrieval for moving object segmentation based on per-pixel mixtures
applications, with emphasis placed on deriving similarity of Gaussians in RGB color space. To improve the robustness

1-4244-0259-X/06/$20.00 2006 IEEE


4851
of the system, a brightness/contrast rate-limiting controller is similarity measure with a curve parameterization that reects
used to lter out sudden and transient changes of illumination the chronological order of the points in a trajectory. Lastly, we
in the input video. In addition, a camera vibration compensa- cannot allow for partial matches between trajectories as this
tion system is used to ensure that the input reects the static may lead to an inconsistent clustering, as shown by [16].
camera assumption made by the tracker. The binarized output
A. Requirements and Motivation
of the background segmentation is processed morphologically,
and connected regions are extracted. Spectral clustering works by considering one or more
The frame-to-frame overlap of connected regions is used eigenvalues of an afnity matrix between the elements being
to compute region associations. Targets are modeled as sets clustered. An ideal similarity measure in that sense is one that
of associated connected regions, and their internal state is produces a block-diagonal afnity matrix [11]. Let s(P, Q)
estimated using an Extended Kalman Filter to reect the be the similarity between for two trajectories P and Q,
non-linearity of the perspective projection. The sequence of represented as ordered sequences of points on the ground plane
Kalman updates itself is recorded in order to allow for the (i.e., P = {pt }0t<Tp and Q = {qt }0t<Tq ). The design goals
trajectories to be rened using Kalman smoothing when a for s(P, Q) that reect the optimality criterion are:
target leaves the trafc scene. A direct geometrical approach 1) s(P, Q) must tend to 1 if P and Q are the trajectories of
based on 3-point perspective is used to t 3D bounding vehicles that passed through the same part of the scene,
boxes to the moving objects outlines. This allows accurate in the same direction.
localization of vehicles on the ground plane and provides 2) s(P, Q) must tend to 0 in other cases.
reasonable estimates of the width, length, and height of the 3) s(P, Q) = s(Q, P ).
tracked objects. The raw trajectory output of the tracker on a B. Modied Hausdorff Distance
T-shaped intersection is shown in Fig. 1.
The Hausdorff distance from the set P to the set Q is dened
For more details on the tracking algorithm, refer to the
as:  
descriptions in [14], [15].
h(P, Q) = max min d(p, q) , (1)
pP qQ

8500 where d(p, q) is any metric. In the case of trajectories, d(p, q)


is taken to be the Euclidean metric. The distance dened by
7500
(1) is a semi-metric, which is usually extended to a metric by
dening H(P, Q) = max(h(P, Q), h(Q, P )), which is known
6500
as the Hausdorff distance between the sets P and Q.
5500
The Hausdorff distance suffers from some well-known de-
ciencies its sensitivity to outliers, and its complete disregard
4500 for the structure of the underlying sets. While the unmodied
Hausdorff distance has been successfully used in template
3500 and curve matching, it is inadequate for the aforementioned
purposes since it fails to account for the direction in which a
2500 trajectory is traversed. Another shortcoming of the distance is
illustrated by the example in Fig. 2.
1500
0 1000 2000 3000 4000 5000 6000

A
Fig. 1. Output from the tracker at a T-shaped intersection (centimeter-scale
grid, 386 trajectories total). The Kalman smoothing results in very clean
trajectories.

B
III. S IMILARITY M EASURE Fig. 2. Two sets with a small Hausdorff distance. The intuitive notion of
Trafc signals, pedestrians and congestion contribute to a similarity between trajectories is not reected by the Hausdorff distance in
this case, since the structure of the sets is ignored.
the large variation in the fashion by which two different vehi-
cles may make their way through a trafc intersection. While We address the task-specic deciencies of the Hausdorff
time-related attributes reveal important information about the distance by imposing a structure on the underlying sets and
activities taking place at a trafc scene, they are irrelevant to by ignoring a portion of the worst-matching points between
the task of determining its physical layout. This suggests the the two sets:
use of a spatial-only similarity measure, and ignoring the time  

dimension of trajectories. However, we may lose the ability h (P, Q) = ord min d(p, q) , (2)
pP qN (C(p))
to discriminate between trafc ow in opposing directions by
completely ignoring the time component of points on a trajec- where N (q) is the neighborhood of q within Q, C(p) is
tory. We address this issue by augmenting an essentially spatial the set of points from Q that correspond to a point p from

4852
P , and ord pP f (p) denotes the value of f (p) that is larger along a trajectory is very small relative to the total length of
than a fraction of all values of f over P (e.g. = 0.5 the collected trajectories even for fast moving vehicles, the
corresponds to the median). Note that the directed Hausdorff approximation is good enough and we forgo any resampling
distance is a special case of (2) since h(P, Q) = h1 (P, Q) for of the trajectories.
any combination of N and C that satises N (C(p)) = Q for
C. Local Scaling
all p P .
A choice of < 1 means that matches between trajectories Once the parameters and w are xed, the denition of the
can be partial, but h (P, Q) will be robust to the presence modied directed Hausdorff distance is complete. We can now
of about (1 )|P | outliers. Applications that require h to use the distance h (P, Q) to dene the similarity between the
be thresholded must account for the fact that h is a non- trajectories P and Q, in the fashion proposed by [11]:
1
decreasing function of . s(P, Q) = e 2 h (P,Q)h (Q,P )/(P Q ) , (5)
The choice of the neighborhood structure N is motivated
by the desire to avoid the situation exemplied by Fig. 2, yet where P and Q are local scale estimates that reect the
allow for a certain degree of slack in the matching of the two density of the neighborhoods of P and Q in the space induced
trajectories. Since we are interested in the spatial properties of by h . The interpretation of (5) is that h (P, Q)/P is the
trajectories, we base our choice of N on the relative distance distance from P to Q viewed on a scale representative of P s
along the trajectory itself and the point of interest q Q: neighbors in trajectory space, and likewise for h (Q, P )/Q .
The resulting similarity measure is symmetric for any h , and
N (q) = {r Q|(r, q)/L(Q) < w}, (3) is equivalent to the traditional denition of afnity based on
2 2
a metric d as ed /2 when h turns out to be symmetric.
where w is a threshold, (r, q) is the length of the portion of
We select the local scale P for a trajectory P automatically
the trajectory Q that lies between the points r and q, and L(Q)
as the distance h (P, R) to P s 9-th nearest neighbor R (as
denotes the length of Q. The measure h is a non-increasing
viewed from P ). The choice of the 9-th neighbor gives good
function of w.
results, but any small integer constant would be appropriate.
The last parameter that needs specication is the correspon-
dence mapping C. The denition of C ensures that we traverse IV. T RAJECTORY C LUSTERING
the two trajectories in chronological order, so that the direction A. Trajectory Preprocessing
of trajectories is taken into account. We dene C(p) to be the
There are two issues that we must address before clustering
point on Q that is at the same relative distance Q from the start
can be applied pedestrian trajectories must be removed
of Q as p is relative to the initial point of P . More precisely,
from the trajectory set reported by the tracker, and a method
C(p) = argmin |(p0 , p)/L(P ) (q0 , q)/L(Q)|. (4) for dealing with outliers (caused by tracking failures) must
qQ be developed. Since we do not need to cluster all trajectories
The shaded region in Fig. 3 illustrates the set of points on Q in order to determine the layout of an intersection, we have
to which the shortest distance from p P is computed. signicant leeway in removing trajectories from consideration,
and we can afford to err on the side of caution when choosing
the input to the clustering method. In the further discussion,
N(C(p))
C(p) let P be a T -point trajectory for an object of dimensions w l
q0 Q as provided by the tracker. The information available to us at
each time instant t 0 . . . T 1 is the position pt (xt , yt ) of
P the object in the ground plane, and its instantaneous velocity
p0 st (ut , vt ).
p
(p0,p) All of the following rules must be satised by trajectories
that are chosen to be clustered. If any of the rules is not
Fig. 3. An illustration of the structural neighborhood N , the correspondence
C and other quantities relevant to the denition of h . satised, we consider the trajectory P undesirable and ignore
it from further consideration:
The denition of h can be extended to continuous rep- T > 90 (6)
resentations of the trajectories if generalized order statistics
(l > w) w [1m, 4m] l [1.5m, 6.5m] (7)
are used, and in fact our denition is an approximation of
the continuous case, with irregular sampling (we sample at max st  [10km/h, 100km/h] (8)
0t<T
xed time intervals, not xed length increments). An argument
pt pt+  2
in favor of irregular sampling can be made based on the max > (9)
0t<t+ <T (pt , pt+ ) 2
observation that slowly moving vehicles can turn with a
smaller turning radius, so more closely spaced points may The rst rule simply rejects trajectories of length 90, which
be necessary to capture the shape of the resulting high- corresponds to 3 seconds given our trackers sampling rate
curvature trajectory (turning radius is inversely proportional of 30 Hz. The rule (7) reects the observation that vehicles
to curvature). Since the distance between consecutive points are invariably longer than wider and provides a range of

4853
acceptable absolute vehicle dimensions. The rule (8) rejects C. The Clustering Algorithm
all objects that never move faster than 10 km/h (which is a Our exposition of the clustering algorithm closely follows
reasonable upper limit on pedestrian walking speed), or objects the description given in [10], which we reproduce here for
that exceed a speed of 100 km/h (which typically happens completeness. This particular spectral clustering method is
as a result of tracking failure). Finally, rule (9) imposes called the Ng-Jordan-Weiss (NJW) algorithm, and we deviate
a maximum curvature constraint on trajectories in order to from it only in the introduction of local scaling in the n n
eliminate problems such as the one displayed in Fig. 4. The afnity matrix A, which was suggested in [11]. The number
parameter is set to a small integer, so that the computed of clusters is xed to be k in advance.
ratio approximates
the true curvature of the path, while the
threshold 2/2 rules out trajectories that make abrupt turns
1) Compute the afnity matrix A according to (11).
of over 90 . 2) Dene D to be a diagonal matrix with Dii = 1jn Aij ,
and compute the normalized afnity matrix as L =
D1/2 AD1/2 .
3300

3) Find x1 , x2 , . . . , xk the eigenvectors of L corresponding


2800
to the k largest eigenvalues. In the case of repeated eigen-
values, ensure that x1 , x2 , . . . , xk are mutually orthogonal.
Form the n k matrix X = (x1 , x2 , . . . , xk ) by concate-
2300
4850 4350 3850 3350 2850 2350 1850
nating the eigenvectors.
4) Normalize each row of X to have unit length, that is, let
Fig. 4. A high-curvature, self intersecting trajectory. Typically caused by 2
Xij Xij / 1jk Xij .
tracking failures of vehicles waiting at a stop light, in the presence of many
occlusions. 5) Cluster each row of the normalized X into k groups using
K-means. A trajectory Pi belongs to cluster j if and only if
the i-th row of X is assigned to cluster j by the K-means
B. Afnity Matrix Computation and Trimming procedure.
6) The ratio of the within-class scatter to between-class scatter
Let {Pi }1im be the trajectories that satisfy the rules of the rows of X is the distortion score that indicates the
established in the previous section. We must rst compute the quality of the clustering.
m m pairwise distance matrix H as:
Hij = h (Pi , Pj ), (10) Fig. 5. The NJW clustering algorithm.

which gives us all information necessary for computing the 9- The automatic choice of the optimal number of clusters is
(9)
th nearest neighbor Ni of each trajectory. We now remove based on minimizing the distortion computed in the last step
(9)
all trajectories Pi for which h (Pi , Ni ) exceeds a certain of the NJW algorithm over a number of possibilities. In our
threshold Tmax , as those trajectories are most likely orphans experiments, we minimized the distortion for values of k from
(since even their 9-th closest neighbor is too far away). Re- 10 to 25.
moved trajectories are put in an articial cluster for purposes
of cluster comparisons (details follow in Section V-A). In all V. E XPERIMENTS
of our experiments, we set Tmax = 3.66m 12ft, which is the We collected video from a major trafc intersection in a
width of lanes on most roads. For the remaining n trajectories, metropolitan area. The video was shot in the early afternoon
we can compute the n n afnity matrix A as: of a sunny day with strong winds (which necessitated the
 video shaking compensation). We processed a 45 minute
0 if i = j
Aij = (11) segment, which resulted in 4004 trajectories total. Note that
s(Pi , Pj ) if i = j,
the number of trajectories is large for two different reasons
where by slight abuse of notation, Pi refers to the trajectories the intersection is very busy and sees signicant pedestrian
not removed by the trimming step. trafc as well, but in addition, our tracker reports trajectories
Trimming of outliers before clustering is important since the of objects that were present in the scene for as few as 10
clustering process uses a number of clusters that is xed in video frames, which can result in many false object detections
advance, which forces an outlier to be assigned to a cluster de- during periods of high camera instability or extensive scene
spite any incompatibilities it may have with its assigned group. clutter. Instead of compensating for such spurious trajectories
One way to combat this problem is to allow for a large number in the tracker itself, we opted for the pre-ltering process
of groups, which results in outliers being assigned to clusters described in Section IV-A. After ltering, 565 highly reliable
with few elements that can be pruned subsequently. This trajectories remained, on which all subsequent experiments
approach proved to be inferior to trimming before clustering, were performed.
since an increase in the number of clusters results in multiple
clusters even for groups of trajectories that are very similar, A. Comparing Clusterings
it increases computation time, and articially decreases the In order to evaluate the sensitivity of the clustering algo-
cluster distortion score which is used to automatically nd the rithm to small changes in its parameters, or to measure its
optimal number of clusters. robustness to noise, we need a way to compare clusterings.

4854
Given two different k-class labelings g1 : [1, n] [1, k] and C. Optimal Parameter Choice
g2 : [1, n] [1, k] of the trajectories {Pi }1in , we build a The automatic choice for the number of clusters minimizes
k k concordance matrix C: the K-means distortion score, so we decided to use the

Cij = (g1 (Pk ) i)(g2 (Pk ) j), (12) distortion score as the objective function to minimize over
1kn various settings for the algorithm. The sensitivity analysis
suggests that we only need to optimize the choice of , and
where is the Kronecker delta. If the labelings g1 and g2 agree
need not focus on w as it affects the clustering process very
perfectly, C will be a (possibly permuted) diagonal matrix. We
little. Our initial experimentation indicated that values of w
reduce the concordance matrix to a single cluster similarity
close to 0.5 and values of larger than 0.8 tend to produce
score s by nding the least squares solution of:
the least distortion in the nal clusters (as measured in the
(c1 , c2 , . . . , ck )T s = (m1 , m2 , . . . , mk )T , (13) last step of the NJW algorithm). With w xed to 0.5, we
 searched for the least distortion by varying from 0.8 to 0.95
where cj = 1ik Cij and mj = max1ik Cij . A higher
in increments of 0.01. The graph in Fig. 8 shows the best
s-value indicates higher similarity between the labelings g1
distortion score obtained for each value of .
and g2 . If g1 and g2 are identical up to a permutation, then
s = 1. x 10
4

B. Sensitivity to Parameter Choice 7

Our rst experiment was to test the sensitivity of the clus- 6

tering procedure to variation in the choice of w and , which 5


directly affects the similarity scores between trajectories. The
4
average similarity between clusterings obtained with different 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.95

parameter values and a reference clustering was obtained, for Fig. 8. Distortion scores for values of from 0.8 to 0.95. The best distortion
score of 4.22 104 is achieved at = 0.88.
values of k ranging from 10 to 25. The average similarity for
clusterings with w ranging from 0.25 to 0.75 is shown in Fig.
6. Similar results when varies from 0.5 to 0.95 are shown D. Robustness to Noise
in Fig. 7.
In order to test the signicance of using Kalman smoothing
1.025 on the trajectories reported by the tracker, we tried adding
1 Gaussian noise of different magnitude to each point in the
0.975 trajectories. By adding noise similar in level to the error
0.95 covariance of the Kalman lter while tracking, we effectively
0.925
0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 cancel the smoothing. As Fig. 9 indicates, the clustering results
Fig. 6. Average concordance of clusterings with held constant at 0.95, at various noise levels barely differ from the clustering with
relative to the clustering with w = 0.5. The parameter w varies from 0.25 no noise added. The only notable change in this process was
to 0.75. that rule (9) rejects more trajectories, as the addition of the
noise increases the local curvature of trajectories.

1
1

0.9
0.995

0.8
0.99
0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95

Fig. 7. Average concordance of clusterings with w held constant at 0.3,


0.985
relative to the clustering with = 0.95. The parameter varies from 0.5 to 0 1 2 3 4 5 6 7 8 9 10

0.95.
Fig. 9. Clustering sensitivity to adding Gaussian noise with standard deviation
ranging from 0cm to 10cm to every point in a trajectory. Concordance is
The choice of w is not critical to the clustering process; relative to the clustering without noise.
as Fig. 6 reveals, the concordance of the obtained clusterings
is better than 0.95 in all cases, which means that more than
95% of the trajectories typically stayed in the same cluster E. Clustering Results
(relative to the clustering with w = 0.5). As Fig. 7 shows, the The optimal clustering recovered with the settings discov-
method is more sensitive to the choice of the clustering ered in the previous sections has 11 groups and is shown in
concordance drops to just below 0.8, and unlike the results Fig. 10. The grouping of the Southbound-Eastbound left-turn
obtained when varying w, the disparity in obtained clusterings lane with the Eastbound trafc lanes is unexpected at rst, but
grows proportionally to the difference between the values can be explained by the fact that the number of left-turning
used. trajectories is very small (only 5 have an initial point detected

4855
during the approach to the turn). When a higher number structures on the sets between which distance is measured, in
of groups is used, those lanes are separated, but the major order to better reect intuitive distinctions between trajectories
Northbound trajectories are split in more than two groups in the ground plane. The natural t between directed distance
(depending on the initial vehicle detection location). Parallel measures and local scaling of similarity values in an afnity
lanes were typically clustered together when a number of lane matrix is exploited within the spectral clustering framework.
changes occur (as is the case with Southbound, Eastbound and The sensitivity of the method to parameter choices and noise is
Westbound trafc). The results are excellent considering that investigated using a large data set obtained automatically from
the scale and number of groups are automatically found. A a challenging video acquired at a busy trafc intersection.
selected subset of the trajectories is shown projected onto the
VII. ACKNOWLEDGMENTS
original input video in Fig. 11.
This work has been supported in part by the National Sci-
ence Foundation through grant #IIS-0219863, the Minnesota
6500
Department of Transportation, and the ITS Institute at the
6000 Univeristy of Minnesota.
5500 R EFERENCES
5000 [1] Z. Kim and J. Malik, Fast vehicle detection with probabilistic feature
grouping and its application to vehicle tracking, in Proc. IEEE Intl
4500
Conf. Computer Vision, vol. 1, 2003, pp. 524531.
4000
[2] D. R. Magee, Tracking multiple vehicles using foreground, background
and motion models, Image and Vision Computing, vol. 22, no. 2, pp.
3500 143155, Feb 2004.
[3] H. Veeraraghavan, O. Masoud, and N. P. Papanikolopoulos, Computer
3000 vision algorithms for intersection monitoring, IEEE Trans. Intelligent
Transportation Systems, vol. 4, pp. 7889, Jun 2003.
2500 [4] J. Badenas, J. M. Sanchiz, and F. Pla, Using temporal integration for
tracking regions in trafc monitoring sequences, in Proc. 15th Intl
2000
Conf. Pattern Recognition, vol. 3, Sep 2000, pp. 11251128.
1500 [5] F. Porikli, Learning object trajectory patterns by spectral clustering,
in IEEE Intl Conf. Multimedia and Expo, vol. 2, June 2004, pp. 1171
1000 1174.
7500 7000 6500 6000 5500 5000 4500 4000 3500 3000 2500 2000 1500 1000
[6] D. Buzan, S. Sclaroff, and G. Kollios, Extraction and clustering of mo-
tion trajectories in video, in Proc. 17th Intl Conf. Pattern Recognition,
vol. 2, Aug 2004, pp. 521524.
Fig. 10. The resulting clustering of the trajectories with = 0.88, w = 0.5. [7] L. Chen, M. T. Ozsu, and V. Oria, Robust and fast similarity search
for moving object trajectories, in Proc. ACM SIGMOD Intl Conf.
Management of Data. New York, NY, USA: ACM Press, 2005, pp.
491502.
[8] M. Vlachos, D. Gunopulos, and G. Das, Rotation invariant distance
measures for trajectories. in Proc. 10th ACM SIGKDD Intl Conf.
Knowledge Discovery and Data Mining, Aug 2004, pp. 707712.
[9] M. Vlachos, D. Gunopoulos, and G. Kollios, Discovering similar mul-
tidimensional trajectories, in Proc. 18th Intl Conf. Data Engineering,
2002, pp. 673684.
[10] A. Y. Ng, M. I. Jordan, and Y. Weiss, On spectral clustering: Analysis
and an algorithm, in Advances in Neural Information Processing
Systems 14, T. G. Dietterich, S. Becker, and Z. Ghahramani, Eds.
Cambridge, MA: MIT Press, 2002, pp. 849856.
[11] L. Zelnik-Manor and P. Perona, Self-tuning spectral clustering, in
Advances in Neural Information Processing Systems 17, L. K. Saul,
Y. Weiss, and L. Bottou, Eds. Cambridge, MA: MIT Press, 2005, pp.
16011608.
[12] D. P. Huttenlocher, G. A. Klanderman, and W. J. Rucklidge, Comparing
images using the Hausdorff distance, IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 15, no. 9, pp. 850863, Sep 1993.
[13] M.-P. Dubuisson and A. K. Jain, A modied Hausdorff distance for
object matching, in Proc. 12th Intl Conf. Pattern Recognition, vol. 1,
Oct 1994, pp. 566568.
[14] S. Atev, O. Masoud, R. Janardan, and N. P. Papanikolopoulos, A
collision prediction system for trafc intersections, in Proc. IEEE/RSJ
Fig. 11. Trajectories from each cluster overlaid onto the input video.
Intl Conf. Intelligent Robots and Systems, Aug 2005, pp. 169174.
[15] S. Atev, H. Arumugam, O. Masoud, R. Janardan, and N. P. Pa-
panikolopoulos, A vision-based approach to collision prediction at
VI. C ONCLUSIONS trafc intersections, IEEE Trans. Intelligent Transportation Systems,
vol. 6, no. 4, pp. 416423, Dec 2005.
We have presented a method for automatically identifying [16] N. Meratnia and R. A. de By, Aggregation and comparison of tra-
important spatial structures of a trafc intersection using tra- jectories, in Proc. 10th ACM Intl Symp. Advances in Geographic
jectories of vehicles. The directed Hausdorff distance is gener- Information Systems, Nov 2002, pp. 4954.
alized by the imposition of neighborhood and correspondence

4856

Das könnte Ihnen auch gefallen