AN EFFICIENT SYSTEM FOR FORWARD COLLISION AVOIDANCE USING LOW COST CAMERA & E...
SZADKOWSKA Joanna 228942 Poster
1. Instrument data acquisition
using a mobile phone camera
Joanna Szadkowska
Supervisor: Dr Jane Hodgkinson
School of Aerospace, Transport
and Manufacturing
Cranfield University
Background
Graphical User Interface
Conclusion
Developed system
MSc Computational and Software Techniques in Engineering 2015 (Digital Signal and Image Processing Option)
While new technologies and devices spread at a tremendous pace there is
constant need to save data on the computers. However some devices do not
have an interface to connect to a computer. Nevertheless, measurement still
can be accessed by their screen with use of eyes or camera and Computer
Vision techniques.
Images / video from
mobile phone camera
Application
processes frames
Result *.txt file with
reading from device
Images sources: http://dakimg.com/detail.php?id=259; https://www.colourbox.com/; http://www.conrad-electronic.co.uk
collect data
with mobile
phone
run
application
choose
settings
enjoy text
file with
measurement data
Automatic Regions Of Interest detection
0
5
10
15
20
25
30
1
21
41
61
81
101
121
141
161
181
201
Noofwhitepxs
column number
Screens detection
Text regions detection
– sliding window
Cropping ROI
Choose
source
Choose
ROIs
Choose
thresholds
Recognize
or train
Thresholds correction enabled Recognition with
Artificial Neural Network
The application is built
with the use of C++
language, employing the
open source OpenCV
library. The Graphical User
Interface uses the QT
library.
Key results
ROI detection accuracy (1812 samples)
Image source video pictures together
Accuracy 93.8% 98.6% 94.2%
Tested with samples taken in laboratory from different positions
(pictures as well as video).
Recognition accuracy (1995 samples)
Samples from pictures (1920 x 2560 px) and videos (720 x 1280 px).
Image source video pictures together
Accuracy 99.6% 99.7% 99.7%
Whole system accuracy (1068 ROIs)
Image source video pictures together
Accuracy
(transitions incl.)
92.3% 95.1% 92.7%
Accuracy
(without transit.)
94.6% 97.9% 95.0%
Proposed system met its goals. The main objective of this
project was implementation of method for automatic ROI
detection. Submitted method proved to be successful on data
randomly selected from the set of videos and pictures
collected in the laboratory taken by a smartphone.
System can be easily adapted to use with other devices.
What causes mistakes?
• Digits during transition;
• Camera wobble – camera should be fix-positioned;
• Dependency on the light of automatic ROI detection. The best situation
is when lightening is uniform and there are not reflections on the screen;
• Low resolution of images / video and too big camera distance from
the screen.
Application is easy to use, intuitive, it prevents mistakes and informs about
illegal actions. User manual included.
Text regions detection based on column histogram (white pixels in every
column are counted), then algorithm detects characteristics for text with
regions – e.g. gaps between characters.
However, to work precisely it needs data recorded in suitable manner
(distance from the screen, illumination, pictures/video resolution).
Description about data collection is included in the user manual.