Download Fast Human Detection in Crowded Scenes by Contour Integration

Fast Human Detection in Crowded
Scenes by Contour Integration and
Local Shape Estimation
Csaba Beleznai, Horst Bischof
Computer Vision and Pattern Recognition, 2009
Outline
• Introduction
• Outline of the detection method
• Shape-based detection
– Contour integration by integral images
– Human detection by sparse contour templates
• Detection using approximated Shape Context
• Detector combination, optimization
• Experiments and discussion
Introduction
• High detection rates and low false alarm rates are
essential.
• Perform for spatially separated, unoccluded
humans well, but worse for high density of
humans.
• Shape-based detection rates drop significantly in
presence of occluded humans.
• Motion-based detection errors become evident
with increasing density of humans and clutter.
Introduction
• Lin et al. [10] propose a hierarchical contour
template matching scheme combined with
motion detection and human inter-occlusion
analysis.
• Need large computation
• This paper combined Contour Integration and
Local Shape Estimation in real time
Outline of the detection method
Contour integration by integral images
• Use discrete unit-integer orientations
Contour integration by integral images
0-45 135-180 degrees
45-135 degrees
Contour integration by integral images
• Scan the image line-by-line
Human detection by sparse contour
templates
• Generating sparse contour templates
– 120 pedestrian images of the INRIA dataset [5]
– using PCA and 11 eigenvectors are retained
explaining 95% of the total variance
– Generate 30 shape sample stemplates
• input image is filtered along the unit-integer
orientations and filter responses are thresholded to
obtain edge probability maps
Human detection by sparse contour
templates
•
denote the locally best matching headshoulder and full-body templates, w1 and w2 are
importance weights.
Detection using approximated Shape
Context
• A small set (10 images)
of manually segmented
binary images.
• 3*3 cell
• Background subtraction
Detection using approximated Shape
Context
where p (x |li ) denotes the learned spatial
distribution of the best matching codebook
entry and p (li |I) = Ct (li | I ) is its likelihood.
Detector combination, optimization
• The two detector outputs are combined in a
similar manner as in [10].
Experiments and discussion
• CAVIAR dataset [1] and two of our datasets
(RS-A) and(RS-B).
• Detecting standing person with less than 50%
occlusion
Experiments and discussion
Experiments and discussion
Experiments and discussion