Welcome to Scribd, the world's digital library. Read, publish, and share books and documents. See more
Download
Standard view
Full view
of .
Look up keyword
Like this
2Activity
0 of .
Results for:
No results containing your search query
P. 1
Real-time Detection of Illegally Parked Vehicles using 1-D Transformation

Real-time Detection of Illegally Parked Vehicles using 1-D Transformation

Ratings: (0)|Views: 178|Likes:
Published by matthewriley123

More info:

Published by: matthewriley123 on Jul 03, 2008
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

05/09/2014

pdf

text

original

 
1
 
Real-time Detection of Illegally Parked VehiclesUsing 1-D Transformation
Jong Taek Lee, M. S. Ryoo, Matthew Riley, and J. K. AggarwalComputer & Vision Research CenterDept. of Electrical & Computer Engineering, The University of Texas at AustinAustin, TX 78712, USA
{jongtaeklee, mryoo, mattriley, aggarwaljk} @mail.utexas.edu
Abstract
With decreasing costs of high quality surveillance systems,human activity detection and tracking has becomeincreasingly practical. Accordingly, automated systemshave been designed for numerous detection tasks, but thetask of detecting illegally parked vehicles has been left largely to the human operators of surveillance systems.We propose a methodology for detecting this event in real-time by applying a novel image projection that reduces thedimensionality of the image data and thus reduces thecomputational complexity of the segmentation and tracking processes. After event detection, we invert thetransformation to recover the original appearance of thevehicle and to allow for further processing that mayrequire the two dimensional data. The proposed algorithmis able to successfully recognize illegally parked vehiclesin real-time in the i-LIDS bag and vehicle detectionchallenge datasets.
1.
 
Introduction
The problem of tracking cars in different scenarios hasbeen studied widely due to its applicability to numerouspractical situations. One significant application is theproblem of detecting cars that have been parked in illegallocations. A system capable of accurate, real-timedetection of this nature would serve to automate andgreatly improve surveillance systems and would be of great value to those monitoring both public and privatelocales.In this paper, we present an algorithm for automateddetection of illegally parked vehicles in real-time. Thecontribution of this paper is the development of an imagetransformation that allows us to perform event detectioncomputation quickly without compromising accuracy andthe construction of the entire real-time system intransformed domain.The proposed algorithm consists of three general steps.First, we perform an image projection that transforms thetwo dimensional data of each No-Parking (NP) zone intorepresentative one dimensional data. This transformationaccomplishes two tasks. First, it reduces the timecomplexity of the segmentation and tracking tasks.Segmentation of the NP zone is reduced from quadratic tolinear time with respect to the length of the images, andtracking is simplified significantly by restricting vehiclemotion to a single dimension. Second, the transformationaccounts for the variation in size of the vehicle as it movestowards or away from the camera. Utilizing knowledge of the camera’s orientation with respect to the NP zones, thetransformation produces a representation of the vehiclewhose size is consistent throughout the frames.After transformation to 1-dimension, the segmentationprocess relies on both background subtraction andconsecutive frame subtraction. Tracking is based onmatching with a cost function. Merging and splittingoperations are performed to account for occlusion of thevehicles. Together, segmentation and tracking allow theautomatic detection of vehicles that are parked illegally.After performing event detection in one dimension, weinvert the transformation to restore the original appearanceof the vehicles, which may be required for furtherprocessing, such as vehicle model classification. Thus, bysimplifying the otherwise computationally expensivecomponents of any vision system, our algorithm can makereal-time decisions, while preserving all originalinformation. This dramatically increases the practicalapplicability of our system.Relevant work is briefly mentioned in section 2. Insection 3, the proposed method is described. 1-Dprojection (section 3.1), segmentation (section 3.2),tracking (section 3.3), and reconstruction (section 3.4) areprovided. We present tracking results with comments insection 4. The paper concludes with section 5.
2.
 
Previous work
There is plenty of literature relating to vehicle detection,but there is a paucity of literature on specialized vehicleactivities, such as detection of illegally parked vehicles orcollisions. However, a patent [1] has been awarded for a
 
2
 
Figure 1
: transformation of curved representative lines of  NP-zones to adjusted straight lines (This figure is best viewed in color.)
system that detects and warns of an illegally parkedvehicle. For simple vehicle tracking, Gupte
et al
[2] track blobs using an association graph to connect objects oncontinuous frames. Song and Nevatia [3] have been able toovercome difficulties arising from occlusions and track vehicles individually. Haritaoglu
et al
[4][5] track objectsusing concepts of region splitting and merging to handleocclusion.
3.
 
Proposed Method
Our proposed method consists of four consecutiveprocesses: 1-D projection, segmentation, tracking, andreconstruction. After one process is completed, the nextstep is allowed to proceed.
3.1 1-D Projection
In this sub-section, we present our algorithm to convert allthe regions in the image that correspond to NP zones intoone-dimensional vectors. Here, the NP zones are theregions where vehicles are not allowed to park, and theyare known for the given video sequences. We usecorresponding masks to extract these zones in each image.Usually, NP zones lie along the sides of roads, so we focuson the fact that most often roads are curved, and thereforeso are the zones of interest. In order to detect and track vehicles in the NP zones more efficiently and reliably, oursystem transforms these zones into our 1-D representation.Using our proposed transformation, vehicles moving at aconstant speed along the direction of the road aretransformed into vehicles moving along a straight line atthe same speed, and they are normalized to have constantsize across every frame.To this end, we first define a representative line foreach NP-zone, which is composed of a set of pixels. Thisset of pixels on the representative line is mapped onto onestraight line, according to its location, as shown in Figure1. In order to map other pixels in the region of the NP-zone onto the straight line, we first locate for each of thoseFigure 2
: an image with grid to show a process to definethe ordered set of pixels u and to find coordinates of pixelson an NP-zone (This figure is best viewed in color.)
pixels the closest pixel on the representative line. Thosepixels are mapped to the same point on the straight line asthe closest pixel in the representative line, we find themean and variance of the RGB color information of allpixels mapped there. This process is explained more indepth as follows.Let
u =
{
u
i
| 1
 
i
 
 
n
} be the ordered set of pixelsrepresenting one edge of an NP zone in the two-dimensional image as shown in Figure 2.Now, before transformation to 1-dimension, we firstdefine how the transformed coordinate of each pixel
u
i
onthe representative (curved) line
u
is decided.
u
istransformed into a new straight line in the form of a 1-Dvector which only has vertical location information. Thisis computed in the following way:
( )
1
0
Coordinateu
=
 
(1)
( ) ( ) ( )
11
,
iiii
CoordinateuCoordinateuDistanceuu
+ +
= +
 
(2)
Coordinate
(
u
i
) is the new coordinate of 
u
i
 
on thestraight line, and
 Distance
(
u
i+1
,
u
i
) represents thenormalized distance between the two adjacent pixels,
u
i+1
 and
u
i
,
 
in
 
the original image, such that
( )( )
( )
( )
11cc1
||,()/
iiiiiicin
uu DistanceuumeanLengthumeanLengthun
++
=
(
3)Due to this appropriate
 Distance
function, the lengthsof transformed representative lines for vehicles in Figure 1
 
3
 
are similar. On the right hand side of Eq. (3), thenumerator |
u
i+1
u
i
| is the distance between pixel u
i+1
andu
i
in the original image, which is distorted by a camera. Tocompensate camera distortion errors, the denominatorshould be proportional to the distance of the point
u
i
fromthe camera. To this end, we track several sample cars(car
c
)through a sequence and find out the length of cars
 Length
c
(
u
i
) for all those cars and all pixels of therepresentative line,
u
i
.
 Length
c
(
u
i
) can be estimated eithermanually or automatically using the information of widthand height of car blobs. With the obtained lengthinformation, we are able to estimate normalized distanceas shown in Eq. (3). The size of objects represented in thetransformed 1-D vector can be normalized. As shown inFigure 3 (a) and (b), the size of all similarly-sized vehiclesat different locations in the original image becomes similaron the projected image.After obtaining the new coordinates on the 1-D vectorline, we transform the whole image region of a NP zoneinto the same vector representation as shown in Figure 2.For each pixel
v
in the image region of an NP zone, wefind the corresponding coordinate on the representativeline from
u
i
, which is the closest pixel to
v
among allpixels on the representative line as shown in Eq. (4).
( ) ( )
( )
argmin,
i
i
uu
CoordinatevCoordinateDistancevu
=
 
(4)Next, the mean and variance of color (in the RGBspace) is calculated along a set of pixels {
v
} such that allvalue {
Coordinate
(
v
)} is same to
Coordinate
(
u
s
), where
u
s
is the closest pixel to the set of pixels {
v
}. As a result,the image region in 2-D space (associated with the NPzone) is converted into a 1-D vector of color values. Theresult is shown in Figure 3 (c).
3.2 Segmentation
Segmentation is based on two methods: backgroundsubtraction and differencing of adjacent frames in atemporal image sequence. The methods we use are thesame as the regular methods used for 2-D images. After 1-D projection, the size of the data we need to handle ismuch smaller.For background initialization, we assume that there areno illegally parked vehicles initially. Any backgroundsubtraction techniques can be adopted but we use amedian filter on the 1-D projection of this initial imagesequence in order to obtain backgrounds. To get a generalbackground for a longer video sequence, the backgroundcan be updated by accumulating information of updatingnon-foreground pixels on the initial 1-D projectionbackground and applying median filter again.After calculating the 1-D background of NP zones, weFigure
 
3
: Result of 1-D projection. (a) the original image(b) intermediate straightened translations of two No-Parking zones (c) two band images displaying thecorresponding mean RGB color information. (This figureis best viewed in color.)
segment the foreground objects in each frame bysubtracting the background from the calculated projectionof that frame. Similarly, we subtract subsequent frames totrack objects as they move through the extracted zone. Bycombining these two results, we obtain a very goodestimate of the 1-D foreground blobs.In post-processing, we remove noise in the foregroundby performing morphological operations such as dilating,erosion and filling holes [6]. Connected componentanalysis is used to label the 1-D foreground. Since we arelabeling the 1-D line not the 2-D area, the time required tolabel the blobs decreases. As a result, we obtain a set of regions in the foreground that are likely to represent thetrue foreground objects as shown in Figure 4. Thus, thetime complexity for segmentation is reduced to O(
n
) in 1-D, whereas the time complexity for segmentation in 2-D isO(
n
2
).

Activity (2)

You've already reviewed this. Edit your review.
1 thousand reads
1 hundred reads

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->