Human Detection

1
Human Detection

• Mikel Rodriguez

2
Organization

• 1. Moving Target Indicator (MTI)
• Background models
• Moving region detection
• Target chip generation
• Results

• 2. Target Classification (Human Detection)
• Target features
• Support vector machines
• Results

Input Frame
Object Detection
Target Chips
Wavelet Features
SVM Classifier
MTI
Classification
3
Moving Target Indicator

• Moving target indicator (MTI) identifies moving
  objects which can be potential targets

4
MTI Motivation

• Becoming increasingly important in military and
  civilian applications
• To minimize human involvement
• Expensive
• Short attention spans
• Computerized monitoring system
• Real-time capability
• 24/7

5
MTI Challenges

• Different sensor modalities
• LADAR, IR, EO
• Targets with different dynamics
• Small targets
• Weather conditions
• Illumination changes, shadows

6
MTI
Input Video
dynamic update
7
Hierarchical Approach to Background Modeling

• Pixel level
• Region level
• Frame level

8
Pixel LevelBackground Features

• Intensity, heat index
• Gradient
• 2D magnitude, orientation

IR
EO
Magnitude
Orientation
9
Pixel Level Background Features

• Intensity, heat index
• Per-pixel mixture of Gaussians.
• Gradient based subtraction
• Gradient feature vector ??m, ?dd

10
Pixel LevelMoving Region Detection

• Mark pixels that are different from the
  background intensity model
• Mark pixels that are different from the
  background gradient model

11
Region LevelFusion of Intensity Gradient
Results

• For each color based region, presence ofedge
  difference pixels at the boundaries is checked.

• Regions with small number of edge difference
  pixel are removed, color model is updated.

12
Frame LevelModel Update

• Performs a high level analysis of the scene
  components

If more gt 50 of the intensity based background
subtracted image becomes foreground.
Frame level processing issues an alert
Intensity based subtraction results are ignored
13
Structure of the MTI Class
14
Results
15
Target Classification

• Classification of objects into two classes
  humans and others, from target chips generated by
  MTI

16
Challenges

• Small size
• Obscured targets
• Background clutter
• Weather conditions

17
Classifier Flow
Negative
Positive
Support Vectors
Testing
Decision
18
Wavelet Based Target Features
19
Feature Extraction

• Apply 2D Wavelet Transform
• Daubechies wavelets
• Apply Inverse 2D Wavelet Transform to each of the
  coefficient matrices individually
• Rescale and vectorize output matrices

20
Why Wavelets?

• Separability among samples
• Humans can be separated from cars and background

Correlation using gray levels
Correlation using gradient mag.
21
Why Wavelets?
Person 11 - DB3 Wavelet Correlation
22
Support Vector Machines (SVM)

• Classification of data into two classes
• N dimensional data.
• Linearly separable
• If not transform data into a higher dimensional
  space
• Find separating N dimensional hyperplane

23
SVMLinear Classifier
hyperplane equation
N dimensional data point xi
Sample distance to hyperplane
24
SVMBest Hyperplane?

• Infinite number of hyperplanes.
• Minimize ri over sample set xi
• Maximize margin ? around hyperplane
• Samples inside the margin are the support vectors

25
SVMTraining Set

• Let ? 1,A training set is a set of tuples
  (x1,y1),(x2,y2),(xm,ym).
• For support vectors inequality becomes equality
• Unknowns are w and b

26
SVMLinear Separability

• Linear programming,
• Separator line in 2D w1xi,1w2xi,2b0.
• Find w1, w2, b such that ? is maximized
• Find w1, w2, b such that ?(w)wTw is minimized

27
SVMSolution

• Has the following form
• Non-zero ?i indicates xi is support vector
• Classifying function is

28
Classification Class
29
Classification Baseline Analysis

• Run time for 3.0GHz dualcore, 2GB RAM
• Training 276 training samples 8.015 seconds
• Testing 24.087 chips (25 by 25) per second
• Classifier size
• Depends on diversity of images
• For 276 training samples of 25x25, classifier
  size is 1.101 MB

30
Classification Baseline Analysis

• Memory requirements
• Requires entire set of support vectors
• Current classifier

31
Experiments

• Vivid Dataset UCF Dataset

32
Results

• Training set
• 300 target chips
• Testing
• 3872 human chips
• 5605 vehicle and background chips
• Performance
• 2.4 false positive (others classified as
  pedestrians)
• 3.2 false negative (pedestrian classified as
  others)

33
Future directions

• MTI
• Detection by parts
• Motion clustering
• Classification
• Various kernels for SVM
• Better target features
• Motion, steerable pyramids, shape features
  (height, width)
• Local wavelet coefficients
• Adaboost