iBAT: Detecting Anomalous Taxi Trajectories
from GPS Traces
Daqing Zhang†, Nan Li‡,, Zhi-Hua Zhou‡, Chao Chen†, Lin Sun†, Shijian Li
† Institut TELECOM SudParis, France
{daqing.zhang, chao.chen, lin.sun}@it-sudparis.eu
‡ National Key Laboratory for Novel Software Technology, Nanjing University, China
{lin, zhouzh}@lamda.nju.edu.cn
 School of Mathematical Sciences, Soochow University, China
 Department of Computer Science, Zhejiang University, China shijianli@zju.edu.cn
ABSTRACT
GPS-equipped taxis can be viewed as pervasive sensors and
the large-scale digital traces produced allow us to reveal many
hidden “facts” about the city dynamics and human behav-
iors. In this paper, we aim to discover anomalous driving
patterns from taxi’s GPS traces, targeting applications like
automatically detecting taxi driving frauds or road network
change in modern cites. To achieve the objective, firstly
we group all the taxi trajectories crossing the same source-
destination cell-pair and represent each taxi trajectory as a
sequence of symbols. Secondly, we propose an Isolation-
Based Anomalous Trajectory (iBAT) detection method and
verify with large scale taxi data that iBAT achieves remark-
able performance (AUC>0.99, over 90% detection rate at
false alarm rate of less than 2%). Finally, we demonstrate
the potential of iBAT in enabling innovative applications by
using it for taxi driving fraud detection and road network
change detection.
Author Keywords
Anomalous trajectory detection, GPS trace, isolation-based
anomaly detection, taxi
ACM Classification Keywords
H.2.8 Database applications: Data mining.
General Terms
Algorithms
INTRODUCTION
With recent advances in sensing, communication, storage
and computing, the digital traces left by people while inter-
acting with cyber-physical spaces are accumulating at an un-
precedented rate. The scale and richness of different digital
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
UbiComp’11, September 17–21, 2011, Beijing, China.
Copyright 2011 ACM 978-1-4503-0630-0/11/09...$10.00.
traces provides us with new opportunities to understand so-
ciety behaviours and community dynamics in different con-
texts, showing great potential to revolutionize the services in
various areas ranging from public safety, urban planning to
transportation management [10, 27].
In modern cities, more and more vehicles, such as taxis, have
been equipped with GPS devices for localization and navi-
gation. Gathering and analyzing the large-scale GPS traces
have provided us a great opportunity to reveal the hidden
“facts” about the city dynamics and human behaviors, en-
abling diverse innovative applications [22, 13, 18, 30, 26, 21,
23, 28, 24]. Recent years have witnessed an increasing in-
terest in trajectory anomaly detection [14, 6, 9], which aims
to detect suspicious moving objects automatically. However,
while several aspects of abnormality of moving objects have
been investigated, there are very few works on discovering
anomalous driving patterns by mining GPS traces with prac-
tical applications examined. In this paper, we intend to mo-
tivate our research on anomalous taxi driving trajectory de-
tection with the following potential applications:
EXAMPLE 1. Many people, mostly tourists, are victims of
taxi driving frauds committed by greedy taxi drivers who
overcharge passengers by deliberately taking unnecessary
detours. To ensure quality taxi services, it is crucial to de-
tect and penalize such frauds. Currently, detecting taxi driv-
ing frauds is often done by experienced staff via manually
checking the GPS trajectories corresponding to the taxi rides,
based on complaints from passengers, but this is costly and
not very effective because many frauds are not even noticed
by passengers. As the traces of driving frauds often signif-
icantly deviate from normal ones, it is possible to automat-
ically detect the anomalous driving trajectories by mining
taxi GPS traces and hence taxi driving frauds.
EXAMPLE 2. Urban road networks often change over time
in developing cities, it is important to update these changes
in the digital map. If this is done manually by digital map
providers, it would be expensive and also difficult to cap-
ture the changes in time. If GPS-equipped taxis are viewed
as moving sensors probing the real-time information about
urban road network, then the taxi traces accumulated in a
new and different area might indicate a sudden road network

SD
t 1
t2
t0
t3
Figure 1. An illustration of taxi trajectories between S and D.
change, i.e. a newly-built or blocked road segment nearby.
Hence, detecting anomalous taxi driving trajectories can be
helpful in identifying road network changes promptly.
Consider the taxi trajectories between two places S (source)
and D (destination) as shown in Fig. 1. Assume that the
three clusters of trajectories between the (S;D) pair are de-
fined as normal ones, then the four trajectories (t
0
, t
1
, t
2
, t
3
)
are considered as anomalies since they are “few” and “differ-
ent” from the normal ones. Detecting driving anomalies is a
non-trivial task because of the following challenging issues.
• First, as shown in Fig. 1, there might be different sets of
normal trajectories between each (S;D) pair and these
trajectory clusters usually have different densities or dis-
tance distributions. If we exploit traditional anomaly de-
tection techniques [14, 6, 9] based on distance or density,
it is hard to choose the parameters and identify all anoma-
lous trajectories.
• Second, multiple normal routes between each (S;D) pair
also mean different driving distances. If we directly model
driving distance for anomalous trajectories detection, it is
not able to discover those anomalies whose driving dis-
tance is close to that of the normal trajectories (like t
3
).
• Third, the road network often changes over time in devel-
oping cities: a new (anomalous) route may become nor-
mal and an old road segment can be blocked. Hence, it
is important to be able to detect an emerging cluster of
anomalous trajectories and incorporate these changes in
the model.
• Finally, some traditional anomaly detection methods often
require that the taxi trajectories be represented as fixed-
length feature vectors. However, the real taxi trajectories
are variable-length sequences of points, thus traditional
methods can not be directly used. If we transform them
into fixed-length feature vectors, spatial information can
be lost. Moreover, GPS traces often suffer from the low-
sampling-rate problem since GPS devices usually send
data at a low and changing frequency.
In this paper, we aim to propose a novel anomalous driv-
ing trajectory detection method which addresses the four
challenges above. Firstly, we extract valid taxi rides from
all the taxi GPS traces, split the city map into grid-cells of
equal size, group all the taxi rides crossing the same source-
destination cell-pair, and augment and represent each taxi
trajectory in each source-destination pair as an ordered se-
quence of traversed cell symbols. In such a way, the prob-
lem of anomalous driving trajectory detection is converted to
that of finding anomalous trajectories from all the trajecto-
ries with the same source-destination cell pair. Secondly, for
all the taxi trajectories between a certain source-destination
cell-pair, we define those trajectories that are “few” and “dif-
ferent” from the normal trajectory clusters as anomalies. In-
stead of profiling the normal trajectories and detecting the
anomalies by employing the similarity or density measure,
this paper proposes an Isolation-Based Anomalous Trajec-
tory (iBAT) detection method which exploits the property
that anomalies are susceptible to a mechanism called isola-
tion [20]. Finally, we perform an empirical evaluation of
iBAT with real-world taxi GPS data and show how the two
applications (i.e., taxi driving fraud detection and road net-
work change detection) can be enabled by using iBAT. In
summary, the main contributions of this paper include:
• We identify a new kind of anomalous trajectory detection
problem based on two motivating applications with taxi
GPS traces. We further propose a series of techniques
to transform the problem of anomalous driving trajectory
detection into an easy-to-solve form: finding anomalous
trajectories from all the trajectories with the same source-
destination cell pair, with each taxi trajectory represented
as a sequence of cell symbols.
• To solve the above mentioned problem, we propose an
Isolation-Based Anomalous Trajectory (iBAT) detection
method which exploits the property that anomalies are
susceptible to a mechanism called isolation. To our best
knowledge, this is the first work applying the isolation
mechanism in the trajectory anomaly detection.
• We evaluate iBAT with real-world GPS traces collected
from 7,600 taxis for one month. It achieves remarkable
detection rate with low processing-time, it also outper-
forms the density-based method as a baseline approach in
terms of AUC (i.e., the Area Under the ROC Curve) [4].
• By using two examples (i.e., taxi driving fraud detection
and road network change detection), we show how inno-
vative applications can be achieved by using iBAT.
RELATED WORK
In this section, we briefly review the related work which
can be grouped into three categories. The first category in-
cludes the work on analyzing or exploiting GPS traces with
research issues other than anomaly detection. For instance,
Patterson, Liao, et al. [22, 18] used GPS traces to infer an
individual’s mode of transportation and daily routine, pro-
viding reminders for persons with mild cognitive disabili-
ties when they, for instance, take the wrong bus; Krumm et
al. [13, 8] showed it is possible to predict the destination and
entire route of a vehicle based on historical GPS traces, and
recently reported further results building routable road net-
works from raw GPS traces [7]. Based on the observation
that taxi drivers are experienced in finding the best route to a
destination, Ziebart et al. [30] and Yuan et al. [26] designed
PROCAB and T-Drive, respectively, providing driving direc-
tion guidance by leveraging taxis’ GPS traces. Zheng et