Ziming Zhang, Yucheng Zhao and Yiwen Wan

dealerdeputyAI and Robotics

Nov 25, 2013 (3 years and 8 months ago)

96 views

Ziming

Zhang,
Yucheng

Zhao and
Yiwen

Wan


Introduction
&Motivation


Problem Statement


Paper Summeries


Discussion and Conclusions


Anomaly is a pattern in the data that
does not conform to the expected
behaviour


Also referred to as outliers, exceptions,
peculiarities, surprise, etc.


outlier

outlier


Vehicle behavior is represented as trajectories



When trajectory does conform to dominant
pattern it is detected as anomaly or outlier



Collective Anomalies

Data Input:
Spatio
-
temperal

trajectories of moving objects

Scene Modeling:


Scene Representation: interest points/path


Learning
Model:unsupervised
/supervised



Activity Analysis: virtual fencing, speed
profiling, path classification,
anomaly
detection, online activity analysis and
object interaction characterization


Accurate and efficient representation of trajectories


Defining
a representative normal
pattern is
challenging


The boundary between normal and outlying
behaviour
is often not precise


Availability
of
labelled
data
for training/validation


Data
might contain noise


Normal
behaviour
keeps evolving

Cláudio

Rosito

Jung
, Member, IEEE,
Luciano

Hennemann
, and
Soraia

Raupp

Musse


IEEE TRANSACTIONS ON CIRCUITS AND
SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 18,
NO. 11, NOVEMBER 2008

Representation

Input Trajectories

Initial Clustering

Cluster
Representation using
4
-
D histogram

Event Detection

(x1,y1)

(x2,y2)

(
xn,yn
)


(x3,y3)


……

F=(x1
-
x2,y1
-
y2,x2
-
x3,y2
-
y3…xn
-
xn
-
1,yn
-
yn
-
1)

Representation

Input Trajectories

Initial Clustering

Cluster
Representation using
4
-
D histogram

Event Detection

Representation

Input Trajectories

Initial Clustering

Cluster
Representation using
4
-
D histogram

Event Detection

Representation

Input Trajectories

Initial Clustering

Cluster
Representation using
4
-
D histogram

Event Detection


trajectories collected from trackers


Offline clustering based on Mixture of Gaussian is
used for path modeling


4
-
D histogram is used to represent spatial and
temporal characteristics of each cluster/path for
further event detection such as drift, shift, entry,
bifurcation, confluence, incoherent local speed,
incoherent local orientation pattern


Two dataset(pedestrian and traffic scenario) are
tested and 20 human observers were used for
accuracy validation: the number of evaluation
that agreed with results from proposed method

Nicolas
Saunier

and
Tarek

Sayed



Reduction of public resources on
detecting traffic collision.


Conflicting causes collisions


Conflicting definition


Two or more vehicles closed enough in time
and space


Trajectory representation


A sequence of {x, y,
v
x
,
v
y
}


HMM (Hidden Markov Model)



HMM (Hidden Markov Model)


Sequence of observation = {walk, shop,
clean}


Compute the probability of observing a
sequence, given a model.


Find the state sequence that maximizes the
probability of the given sequence, when the
model is known. (
Viterbi
)


Induce the HMM that maximizes the
probability of the given sequence. (Baum
-
Welch)




K
-
Means clustering



HMM
-
based K
-
means clustering


A set of vehicle trajectories (sequences)


A set of initial HMM (k HMMs)


Step1: Calculating all the probabilities


Step2: Associating the trajectory with HMM


that maximizes probability of the




trajectory


Step3: Updating HMMs based on the




temporary clustering result


Step4: Repeating step 1, 2 and 3 until




convergence has been reached



Training and testing the model


Several instances of conflicting trajectory
pairs train the model to identify mutual
conflicting trajectory clusters


New trajectories are associated with certain
trajectory cluster based on the specific HMM
probability maximization


Conflicting trajectories are identified by their
clustering result.

C.
Piciarelli

*, G.L.
Foresti

Department of Mathematics and Computer
Science, University of Udine, Via
delle

Scienze

206, 33100 Udine, Italy

Available online 21 April 2006


Problem: Classical two
-
step clustering algorithm can
not update cluster dynamically


Solution: On
-
line trajectory clustering approach with a
tree
-
like structure


Goal:Suit for video surveillances sysytems from image
analysis to behavior analysis to detect anomalous
events


Traditional trajectory clustering not suited for detect
anomalous events



off
-
line: not useful in activity analysis



video system: complex structure,from moving
ojects(low level) to behaviour analysis (high)











Representing trajectories as a tree of cluster


Trajectory(Ti): represented by a list of vectors
Tij(representing a spatial position at time j)



Clusters(Ci): organized in a tree
-
like structure that,
augmented with probability information,
represented as a list of vectors


Define a distance or similarit to check if a Ti
matches a given Ci( dynamically), when a Ti
matches a Ci, cluster needs to be updated.



Tree creation steps:


1)building,create tree of clusters from acquried data
dynamically,without waitting the end of trajectory.


2)maintenance as below:


For behaviour analysis, we define that an anomaly is
an event happening rarely. Also we assume that
dangerous events are generally anomalous. An
anomalous trjectory can be defined as a trajectory
matches a path in the tree with low probability.
With probabilitic information, we can implement
anomaly detection.

Papers

Learning

Fashion


Path

modeling


Activity
Analysis

P1

Offline



unsupervised

clustering and 4
-
D
descriptor

7

abnormal
events
detection

P2

Offline

Semi
-
supervised
HMM
-
Based
clustering

Conflicting
traffic

P3

Online

Tree
-
structure

anomaly
detection


Availability of labelled data for
training/validation is not easy and
unsupervised clustering is
favored



online clustering is very important since
normal behaviour keeps evolving


Approaches robust to noisy
trajectories from tracking is preferred






Questions?