Over-head Person Counter

Download Report

Transcript Over-head Person Counter

Over-head Person Counter in
MATLAB
ECE172A
Benny Wong
Research Problem & Motivation
 Overcrowding is quickly becoming an issue in
today’s ever-expanding population. Needed a
method to count people crossing a controlled area.
(ie to prevent room/floor capacity over limit, or
simply the owners wanted to know how many
occupants are within the premises)
– Can be used in controlled areas to count incoming
people
 Library entrances
 Airport Terminals
 Shopping stores
Related Research

People tracking using robust motion detection and estimation
– Background subtract
– Motion band-pass filter to reduce interlaced video noise
– Semi-static Threshold (increases until subtracted image was essentially 0’s in
training period before changing into a non-updating variable

Tracking and counting pedestrians in real-time using a single camera
– Difference images (subtraction)
– Single step of Erosion and Dilation to remove noise
– Uses blobs that can change to track pedestrians when they cross each other
(rectangular patch that has dynamic behavior dependent on location) somewhat
keeps blobs from combining - rectangles need to overlap completely to combine.

Object detection and counting
– Subtraction
– User defined pattern matching size

Grid based template matching
– Matches people with different images dependent on their location on a video grid
Approach
 Chosen methods
– Background subtraction
 Used by every related research or something to the same effect
– Image Closing operation (defined by trials)
 Closes the holes within blob
– Processes using datasets or position-controlled matching wouldn’t necessarily need
solid blobs
– Image Opening operation
 Somewhat truncates limbs for the analysis step, as well as clears noise.
– Band-pass filtering is not needed due to low motion noise
– Region labeling
 Counts objects in video frame
– Length/Width based pedestrian filter (removes non-persons from count)
 Slightly related to rectangular patches method; over-head view eliminates need
for detection to account for depth
Results
 Results showed robust detection of persons in ideal conditions (100%)
 ideal conditions consist of:
– People are walking at least half a person width/length away from each
other
– People are wearing bright or light colors causing a better threshold yield of
their body size
 Problems arise mostly when:
– People are too close – forming one blob that is filtered from the person
count because they fall outside of the limits that determine a region is a
person
– People are too small and wear dark clothes – static threshold further
shrinks their size before analyzing step
– People being too big would also fall outside of limits
 Most problems have been minimized due to my choice of constraining
limits. My constraints were obtained on a trial-basis using two people of
vastly different body size (petite female, and relatively-large built male).
Improvements
 Not use MATLAB, due to slow processing
 Make it Real-Time
 A better range of seeding person-sizes to further
improve limits
 Training period to improve threshold and
determine a background for subtraction.
 Use Blob-based tracking/detection, tracking is
more accurate than counting every frame. (ie zero
persons detected when two people are too close)
Lessons and Milestones
 Attain final working datasets earlier
– Had to work with very low-budget dataset early on
 Ask more questions
– Would’ve sped up the algorithm-choosing process
 Get a better laptop or work on the same platform
– Program ran fine on a desktop, but needed to be limited
when ran on laptop.
 Use a widely support AVI codec
– Ran into codec problems when working on different
computers that didn’t have DIVX video codec.