CS55 Tianfan Xue Adviser: Bo Zhang, Jianmin Li
Outline Introduction Original Algorithm Improved Algorithm System Design & Data Set Performance Evaluation Work Next Step
Introduction Automatically Video Surveillance Human Tracking What is human tracking Why do human tracking Presumption Person is standing & Normal Pose
Original Algorithm Algorithm Design General Framework Probability Evaluation HOG feature Initial Detect Motion Prediction Drawback
Original Algorithm General Framework Frame n State n-1 Predicted State n Human Detector (HOG) State n Motion prediction & Gauss Diffusion Position & Size HOG features validation Training Set Machine learning Offline Online
Original Algorithm Probability Evaluation Definition x t : State in time t z t : Image in time t Z t : Whole image sequence till time t Probability: Gauss Model + Motion Predict HOG output Simplified in Particle Filter
Original Algorithm Initial Detect Randomly Choose 2000 positions in an image Motion Prediction Linear Regression of recent 10 frame Offline Detector HOG features original Edge mapHOG SVM
Original Algorithm Drawbacks Fail to find a person at emergence Detection Rate ↔ Computational Complexity Loss track when partially Occlusion 2-Magnet Effect
Original Algorithm Drawbacks Fail to find a person at emergence Loss track when partially Occlusion 2-Magnet Effect
Original Algorithm Drawbacks Fail to find a person at emergence Loss track when partially Occlusion 2-Magnet Effect When person A (more obvious) pass person B(less obvious), A will attract B’s window
Improved Algorithm 3 Improvement Use salience to cut search space Combine offline-online classifier(online: Color features) Part Detector Problems
Improved Algorithm Using Salience To Cut Search Space Idea: The position people more like emerge (Salience) Method: Detect at only at position with great variance
Improved Algorithm Combine offline-online classifier(online: Color features) Frame n State n-1 Color detect result Predicted State n HOG Classifier Final result Motion prediction & Gauss Diffusion Size & position Color features validation HOG features validation Color Classifier Training Set Machine learning Offline Online
Improved System Part 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 System Part Detector 2 Leg Color Model Not Visible Torso Color Model Visible HS Color Model Visible Torso HOG Model HS HOG Model Final Property
Improved System Problems 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 Design Tracking System XML Debugging output GUI
Data Set Training Data INRIA Person Data Set 2416 Positive Examples, 1218 Negative Examples Testing Data PETS2004(CAVIAR)
Experiment Result Evaluation Compare ground truth windows with detected windows Overlap:(T=0.5) Tracker Detection Rate(TRDR) & False Alarm Rate(FAR) TP: True Positive, FP: False Positive, FN: False Negative
Experiment Result Baseline: With Color Model, With Salience Detect Test1 Use Salience to Detect New Person Random Select Detect Pos Select At Salience Time15.9s/frame4.5s/frame TRDR61.1%66.8% FAR21.9%15.6% Test2 Color Model Without Color Model With Color Model Time2.2s/frame4.5s/frame TRDR9.8%66.8% FAR20.4%15.6%
Work Next Step Improve online-offline classifier How to learn a good color model How to decide a person is disappeared Make a more wide-arrange evaluation
Q & A
Probability Evaluation Bayesian result Particle Filter Space Too Large!!!
2-Magnet Effect Solve 2-Magnet Effect But it will bring some new problems… Gauss Model + Motion Predict HOG output Punishment for 2 close windows No Color No overlap term No Color Overlap term Color No overlap term Color overlap term TRDR46.9%9.8%66.8%9.8% FAR42.1%20.4%15.6%20.0%
Color Model Features: 72-dim HSV histogram Probability Evaluation: Inner Product of 2 feature vectors
Detect Result Performance of other algorithm (Here, different evaluation standard was used) TRDRFAR Our Method56.1%29.4% BBS42.5%72.4% W411.7%92.1% SGM42.8%54.0% MGM38.2%63.3% LOTS47.9%40.3% Track44.4%35.2%