CS55 Tianfan Xue 2005011371 Adviser: Bo Zhang, Jianmin Li
Jan 20, 2016
CS55 Tianfan Xue 2005011371Adviser: Bo Zhang, Jianmin Li
OutlineIntroductionOriginal AlgorithmImproved AlgorithmSystem Design & Data SetPerformance EvaluationWork Next Step
IntroductionAutomatically Video SurveillanceHuman Tracking
What is human trackingWhy do human tracking
PresumptionPerson is standing & Normal
Pose
Original AlgorithmAlgorithm Design
General FrameworkProbability EvaluationHOG featureInitial DetectMotion Prediction
Drawback
Original AlgorithmGeneral Framework
Frame n
State n-1Predicted State n
HumanDetector(HOG)
State n
Motion prediction & Gauss Diffusion
Position & Size
HOG features validation
Training Set Machine learning
Offline
Online
Original AlgorithmProbability Evaluation
Definitionxt : State in time t
zt : Image in time t Zt : Whole image sequence till time t
Probability:1 1 1 1( | ) ( | ) ( | ) ( | )t t t t t t t t tp x Z p z x p x x p x Z dx
1( | ) ( | ) ( | )i i i i it t t t t tx Z p x x p z x
Gauss Model + Motion Predict
HOG output
Simplified in Particle Filter
1
Ni i
t ti
x x
Original AlgorithmInitial Detect
Randomly Choose 2000 positions in an imageMotion Prediction
Linear Regression of recent 10 frameOffline Detector
HOG features
original Edge map
HOG
SVM
Original AlgorithmDrawbacks
Fail to find a person at emergence Detection Rate ↔
Computational ComplexityLoss track when partially Occlusion2-Magnet Effect
Original AlgorithmDrawbacks
Fail to find a person at emergence
Loss track when partially Occlusion
2-Magnet Effect
Original AlgorithmDrawbacks
Fail to find a person at emergence
Loss track when partially Occlusion
2-Magnet EffectWhen person A (more obvious) pass person B(less obvious), A will attract B’s window
Improved Algorithm3 Improvement
Use salience to cut search spaceCombine offline-online classifier(online: Color features)Part Detector
Problems
Improved AlgorithmUsing Salience To Cut
Search SpaceIdea:
The position people more like emerge (Salience)
Method:Detect at only at position with great variance
Improved AlgorithmCombine offline-online classifier(online: Color features)
Frame n
State n-1Color detect result
Predicted State n
HOGClassifier
Final result
Motion prediction & Gauss Diffusion
Size & position
Color features validation
HOG features validation
ColorClassifier
Training Set
Machine learning
Offline
Online
Improved SystemPart Detector (CVPR05’s, Bo Wu)
7%
32%
49%
93%
20%
64%
10%
24%
46%
82%
21%
77%
12.5% 87.5%
34% 65%
31% 68%
HS
Torso
Leg
HS
Torso
Leg
Color Part
Whole
27% 63%
Improved SystemPart Detector 2
LegColor Model
Not Visible
TorsoColor Model
Visible
HSColor Model
Visible
TorsoHOG
Model
HSHOG
ModelFinal Property
Improved SystemProblems
Color model also learns the occlusion object→ Always Output that all parts is visible
When a person disappear, the corresponding detect window still exists
System DesignTracking SystemXML Debugging outputGUI
Data SetTraining Data
INRIA Person Data Set2416 Positive Examples, 1218 Negative Examples
Testing DataPETS2004(CAVIAR)
Experiment ResultEvaluation
Compare ground truth windows with detected windowsOverlap:(T=0.5)
Tracker Detection Rate(TRDR) & False Alarm Rate(FAR)
| |2*| | | |
obs truth
obs truth
A AT
A A
TPTRDR
TP FN
FP
FARTP FP
TP: True Positive, FP: False Positive, FN: False Negative
Experiment ResultBaseline: With Color Model, With Salience DetectTest1 Use Salience to Detect New Person
Random Select Detect
Pos
Select At Salience
Time 15.9s/frame 4.5s/frame
TRDR 61.1% 66.8%
FAR 21.9% 15.6%Test2 Color ModelWithout Color
ModelWith Color
Model
Time 2.2s/frame 4.5s/frame
TRDR 9.8% 66.8%
FAR 20.4% 15.6%
Work Next StepImprove online-offline classifier
How to learn a good color modelHow to decide a person is disappeared
Make a more wide-arrange evaluation
Q & A
Probability EvaluationBayesian result
Particle Filter1 1 1 1( | ) ( | ) ( | ) ( | )t t t t t t t t tp x Z p z x p x x p x Z dx
1( | ) ( | ) ( | )t t t t t tx Z p z x p x x Space Too Large!!!
2-Magnet EffectSolve 2-Magnet Effect
But it will bring some new problems…
1( | ) ( | ) ( | ) ( )t i t t t t overlapx Z p x x p z x p x
Gauss Model + Motion Predict
HOG output Punishment for 2 close windows
No ColorNo overlap
term
No ColorOverlap term
ColorNo overlap
term
Coloroverlap term
TRDR 46.9% 9.8% 66.8% 9.8%
FAR 42.1% 20.4% 15.6% 20.0%
Color ModelFeatures:
72-dim HSV histogramProbability Evaluation:
Inner Product of 2 feature vectors
Detect ResultPerformance of other algorithm (Here, different
evaluation standard was used)TRDR FAR
Our Method 56.1% 29.4%
BBS 42.5% 72.4%
W4 11.7% 92.1%
SGM 42.8% 54.0%
MGM 38.2% 63.3%
LOTS 47.9% 40.3%
Track 44.4% 35.2%