Dressed Human Modeling

advertisement
Object Detect
报 告 人:
日
期:
郭立君
2007年11月
Contents
•
•
•
•
Introduction
Rigid Object Detect
Human Detect Introduction
Histograms of Oriented Gradients for
Human Detection
• Dressed Human Modeling,Detection
• Conclution
Introduction
• Some Background on Object Detection
As mentioned earlier an object detector can be
viewed as a combination of an image feature
set and a detection algorithm.
1.The image descriptors or feature vectors that
they use
2.The detection framework that is built over
these descriptors
Introduction
• Descriptors or Feature Vectors
Sparse Local Representations
Point Detectors (SIFT etc)
Part or Limb Detectors
Dense Representation of Image Regions
Regions and Fragments Based on Image
Intensity
Edge and Gradient Based Detectors
Wavelet Based Detectors
Introduction
• Classfication Method
Generative Approaches: Typically generative approaches
use Bayesian graphical models with Expectation-Maximisation
(EM) to characterise these parts and to model their cooccurrences. Such as Bayesian and Graphical Models
Discriminative Approaches: Discriminative approaches use
machine learning techniques to classify each feature vector as
belonging to the object or not. Such as Support Vector
Machine (SVM) Classifiers and Cascaded AdaBoost
Rigid Object Detection
• Roberts’ Method
Rigid Object Detection
• Model Based Vehicle Motion(Tieniu Tan)
Human Detect Introduction
Challenges
• Wide variety of articulated poses
• Variable appearance/clothing
• Complex backgrounds
• Unconstrained illumination
• Occlusions, different scales
Human Detect Introduction
Applications
• Pedestrian detection for smart cars
• Film & media analysis
• Visual surveillance
• Mobile robot navigation
• Human motion capture
Note
Two approaches to object detection
One approach is to search the whole image
at multi-scales for objects. This is a time
consuming procedure and may result in
multiple responses from a single object.
Another approach is to first segment foreground objects from the background,then
classify each segmented object as human or
non-human.
Dressed Human Modeling,Detection
--Liang Zhao CMU-RI-TR-01-19
• Backgroud
Dressed Human Modeling,Detection
• Backgroud
Dressed Human Modeling,Detection
• Goal
(a)initial contour detection (b) body parts identification
(c) contour prediction
(d) contour alignment
Dressed Human Modeling,Detection
• Goal
focuses on how to classify a previously segmented object
as human or non-human
• Idea
Rrecognition scheme is based on the shapes of body parts
and the relationships between them. Then the questions
left are how to decompose a silhouette into parts, and how
to represent the shapes and the relationships between the
parts
Dressed Human Modeling
• Requirements for a Good Object Class Model
1. It should not depend on scale, orientation, and position of
objects;
2. It should handle view-dependent shape variation;
3. It should be robust to shape distortions resulting from
digitization noise and foreground/background
segmentation errors;
4. It should be robust to partial occlusions of an object;
5. It should allow for articulated moving parts;
6. It should not be influenced by the shape variations
allowed within the class;
7. It should support efficient shape recognition/classification.
Dressed Human Modeling
• Shape Decomposition for Part-Based Representation
Dressed Human Modeling
• Shape Decomposition for Part-Based Representation
Dressed Human Modeling
• Shape Decomposition for Part-Based Representation
Dressed Human Modeling
• Shape Decomposition for Part-Based Representation
Problems and Solutions:
(a) Smooth the boundary of a silhouette.
(b) Remove noise or small local deformations
(c) Exploit high level information(RCR)
Dressed Human Modeling
• Human body model
TRS-invariant probabilistic model:
For the purpose of human detection and model learning, a
TRS-invariant representation of the shapes of parts and the
relationships between them is developed. For the purpose of
modeling the shape variations between individuals and due
to viewpoint changes, probability distributions are employed
to encode the variations of the model parameters.
Dressed Human Modeling
• Human body model
Dressed Human Modeling
• Human body model
A body part is parameterized with a vector (a; l;
x; y;Ө), where a = w/l is the aspect ratio that
captures the general shape of a ribbon, (x; y)
are the coordinates of the origin in the
coordinate frame of the parent part, and Ө is
the intersection angle between the major axes
of this part and its parent part.
Dressed Human Modeling
• Human body model
Three TRS-invariant matrices: A; S; U,
Dressed Human Modeling
• Human body model
TRS-invariant probabilistic model:
Dressed Human Modeling
• Dressed Human Modeling
Dressed Human Modeling
• Dressed Human Modeling
1.Merged body parts(dynamic)
2.An evaluation function( Bayesian similarity
measure-Shape Similarity Measure)
3.A coarse to fine procedure( RCR algorithmrecursive context reasoning)
Dressed Human Modeling
• Dressed Human Modeling
TRS
BSM
Dressed Human Modeling
• Bayesian Similarity Measure and Body Part
Identification
Problem Formulation:
Dressed Human Modeling
• Bayesian Similarity Measure and Body Part Identification
Problem Formulation:
Dressed Human Modeling
• Bayesian Similarity Measure for Human Detection
Decision Rule:
Dressed Human Modeling
• Bayesian Similarity Measure for Human Detection
Resule:
Histograms of Oriented Gradients
for Human Detection
Background
INRIA Rhˆone-Alpes
1. Focus on building robust feature sets
2. Classifier is just linear SVM on
normalized image windows, is reliable & fast
3. Moving window based detector with nonmaximum suppression over scale-space
Histograms of Oriented Gradients
for Human Detection
Background
Histograms of Oriented Gradients
for Human Detection
Feature Sets
Histograms of Oriented Gradients
for Human Detection
Processing Chain
Histograms of Oriented Gradients
for Human Detection
HOG Descriptors
Histograms of Oriented Gradients
for Human Detection
Performance
Histograms of Oriented Gradients
for Human Detection
Descriptor Cues
Histograms of Oriented Gradients
for Human Detection
Conclusions
Histograms of Oriented Gradients
for Human Detection
Demo
the question should be:
y = arg max h(a1,a2,...,an)
for example: if ai maximizes h, y = ai
Download