Moving Vehicle Detection from a Video, CCTV Footage etc. by Image Processing. The algorithm and the steps to be followed for detection is described in the presentation.
WSO2Con2024 - Simplified Integration: Unveiling the Latest Features in WSO2 L...
Moving vehicle detection from Videos
1. MOVING VEHICLE DETECTION
GUIDE PROJECT
MEMBERS
DR. KHELCHANDRA THONGAM
PRIYANK BHARDWAJ
ROHIT KUMAR
SAHU
AJEYA
SIDDHARTHA
2. INTRODUCTION
INTELLIGENT TRANSPORTATION SYSTEM
• Traffic data may come from different sensors such as loop
detectors, ultrasonic
sensors, or cameras.
• Video-based camera systems coupled with computer vision
techniques offers alternative approach to spot sensors.
• VB camera systems are more sophisticated and powerful.
3. MOTIVATION
• There are many published works for moving vehicle detection
but there were certain areas of improvement in each like:
• how to detect the moving vehicles at night, casting shadow
elimination.
• Different from previous works, this algorithm learns from the
known examples and does not rely on the prior model of
vehicles, lighting, shadows, or headlights as example based
classification system is prevalently used in image-based
classification.
4. CHALLENGES
• Spot sensors have limited capabilities and are often both costly
and disruptive to install.
• Main challenge come from cast shadows, vehicle headlights,
and noise which produce incorrect segments.
• Sunlight cause shadows which are difficult to be distinguished
from the vehicle and cause segmentation errors.
• Vehicle headlights and bad illumination cause strong noise.
5. ALGORITHM DESCRIPTION
The whole algorithm includes the following steps:
1. Background estimation
2. block division
3. candidates’ selection
4. Feature extraction,
5. SVM-based classification,
6. Shape representation.
IMAGE
SEQUENC
E
BACKGROUN
D
ESTIMATION
BLOCK
DIVISION
FEATURE
EXTRACTIO
N
SVM BASED
CLASSIFICATI
ON
SVM-
TRAINING
SHAPE
REPRESENTATI
ON
CANDIDA
TE
SELECTIO
N
Fig. referred from-IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 1, JANUARY
2007
6. BACKGROUND ESTIMATION
• In order to detect moving vehicles, we need to estimate the background of
the scene first.
• We will propose improved adaptive background extraction algorithm based
on Kalman filtering.
• The background on the pixel p of the (n + 1)th frame can be defined as
B(n + 1, p) = B(n, p) + β(n, p) + μ(n, p).
• The input image intensity is described as I(n, p) = B(n, p) + η(n, p).
• By combining (1) with (2), we have I(n + 1, p) = B(n, p) + ω(n + 1, p).
Updated area
Sliding window
Image
Fig. referred from-IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 1, JANUARY
2007
7. BLOCK DIVISION AND CANDIDATES SELECTION
• The image will be divided into non-overlapped blocks, and
each block has the same size in a same image.
• We will find out the blocks with a gray-level change.
• The current image will be subtracted from the background to
get the difference image, and we compute its mean value for
each block.
• These candidates include real vehicles, casting shadows,
headlights, or noise.
9. • Histogram is used for the feature extraction.
• The range of the image’s grey level is taken [0,T] where T
=255.
• For difference image d{i,j} the range of its grey scale is
taken between [-T,T] or[0,2T].
• The histogram ha(r) is computed from the image d{i,j}.
• The histogram hb(r) is computed from the edge map E(i,j)
of the image d{I,j}.
• Here r is between –T<=r<=T.
Fig. referred from-wavemetrics.com
10. FORMULAS
• To compute the edge map,the formula =
• E(i, j) = 1/ 2 {|D(i + 1, j + 1) − D(i, j)| + |D(i + 1, j) −
D(i, j + 1)|}.
• A new histogram hc(r) is from by combining the
histogram of the difference image and its edge map.
• The new histogram hc(r) of dimension3T+R is
computed as
• hc(r) = ha(r), 0 ≤ r ≤ 2T hb(r − 2T − 1), 2T + 1 ≤ r ≤
3T + 2.
11. NORMALIZATIO
N
• .Normalisation of the feature is done to adopt the different block division by h(r)=hc(r)/S
where S is the size of block.
• .PCA, optimal orthonormal decomposition will be applied to reduce the noises from
feature.
• .From the collected sample ,a scatter matrix is obtained as.
S= 1/𝑁 𝑖=0
𝑁 ℎ𝑖 − 𝑚 (ℎ𝑖 − 𝑚) 𝑡
,here N =no. of training set.
• hi is the ith vector and (hi − m)t is the transpose of the vector of (hi − m).
• .For S ∈ R(3T +2)×(3T +2), it is easy to obtain its eigenvalues A = diag[λ1,...,λ3T +2], λ1
≥···≥ λ3T +2 and its eigen-vectors V = [ν1,...,ν3T +2]. Then, a new compressed vector
with a dimension p can be computed from the original vector h(r).
• gp = [ν1,...,νp] t h, when p ≤ 3T + 2.
12. EXTRACTED FEATURE
• Using compressed vectors as features, our
experiments will show that the classification result
is better than directly using the original histograms.
Moreover, the computation cost can also be
reduced.
Fig. referred from-
images.google.com
14. SVM
• Support vector machines (SVMs) are a set of supervised
learning methods used
for classification, regression and outliers detection.
• It is used to generalization capabilities which are
achieved through a minimization of bound of the
general error.
• The aim is to define a hyperplane that divide the set of
examples such that all points with the same bound are
on the same side of the hyperplane.
• The extracted features are put into the SVMs for
training and the parameter of the SVM-based classifier
can be obtained.
Fig. referred from-
wikipedia.org
15. DETECTION STAGE
• The input is divided into many blocks.
• The feature vector is generated by PCA from the two kinds of histogram of the
images.
• It is put into the well-trained classifier to judge whether this region is covered with a
vehicle or others (such as shadow, headlight, or noise).
17. SHAPE OF VECHICLES NEEDS TO BE EXTRACTED
FOR VEHICLE COUNTING,SPEED
ESTIMATION,TRACKING ETC.
SVM BASED CLASSIFIER GROUPS BLOCKS INTO
VEHICLES AND NON VECHICLES
BLOCKS CLASSIFIED AS VEHICLES ARE GROUPED
BY THEIR 8 CONNECTIVITY
A CONVEX POLYGON IS DERIVED FROM A SET OF
RELATED BLOCKS
THE POLYGON IS USED TO APPROXIMATE A
VECHICLE’S SHAPE IN THE FORM OF A
PARALLELOGRAM WHICH IS MORE ROBUST DUE
TO ITS SIMPLE REPRESENTATION
Fig: 8 direction
connectivity
Fig:Parallelogram
P1P2P3P4Fig. referred from-IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 1, JANUARY 2007
18. Fig: White Boxes indicate
detected parts of vehicles
Fig: Shape representation
Fig. referred from-IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 1, JANUARY
20. PROJECT’S TIME DIVISION
S.No. Description Date
1. Research 10/10/17 – 11/11/17
2. Project Implementation 15/11/17 – 31/01/18
3. Documentation 01/02/18 – 25/02/18
21. FUTURE WORK
THE ALGORITHM CAN BE MADE MORE PRECISE BY
USING MORE INFORMATION SUCH AS COLOR
USING SOME OTHER CLASSIFIERS SUCH AS CNN OR A
COMBINATION OF CLASSIFIERS TO IMPROVE
ACCURACY
22. REFERENCES
JIE ZHOU , DASHAN GAO AND DAVID ZHANG , “MOVING
VEHICLE DETECTION FOR AUTOMATIC VEHICLE
MONITORING,” IEEE TRANSACTION ON VEHICULAR
TECHNOLOGY , VOL 56 , NO. 1 , PP. 51-59, JAN 2007
ALI SHARIF RAZAVIAN ,HOSSEIN AZIZPOUR, JOSEPHINE
SULLIVAN, STEFAN CARLSSON, “CNN FEATURES OFF-
THE-SHELF: AN ASTOUNDING BASELINE FOR
RECOGNITION”, 2014 IEEE CONFERENCE ON
COMPUTER VISION AND PATTERN RECOGNITION
WORKSHOPS