# E K O R T S N S

• Slides: 18

E K O R T S N S IO I T N C N E E T ET 5 6 D S 3 C Mentor: Prof. Amitabh Mukherjee Khushdeep Singh (10351) Aakash Verma (10002)

APPLICATIONS OF HUMAN ACTIVITY RECOGNITION 1 1 2 20 03 2 - KHUSHDEEP SINGH AND AAKASH VERMA 2

E B E N H A C T O K T S N A I T D NG I R E U ID W S O IV O SK L D L TA FO UB 1. S 1 1 2 20 03 2 - Player Tracking 2. Optical Flow Analysis 3. Using Motion Descriptors 4. Adaptive Boosting KHUSHDEEP SINGH AND AAKASH VERMA 3

T R L G I E F NLE I Y K A IC T C L R P A PA R T ING U S E R

WHAT MUST BE SPECIFIED: • Prior Distribution: p(x 0) - Describes initial distribution of object states • Transition Model: p(xt | xt-1) - Specifies how objects move between frames - A simple model: sample next state from a Gaussian window around current state - We used second order auto regressive model. xt = Axt-1 + Bxt-2 +wt • Observation Model: p( yt | xt ) - Color Histogram Object Tracking and Particle Filtering by Rob Hess [2006] 1 1 2 20 03 2 - KHUSHDEEP SINGH AND AAKASH VERMA 5

EXAMPLE OF PLAYER TRACKING 1 1 2 20 03 2 - KHUSHDEEP SINGH AND AAKASH VERMA 6

W O M L F TH R L A GO I L C A I L T K P OS IN G U 1 1 2 20 03 2 - KHUSHDEEP SINGH AND AAKASH VERMA 7

OPTICAL FLOW ESTIMATION SPARSE OPTICAL FLOW DENSE OPTICAL FLOW • Computed only at a subset of image points. • Computed at each image pixel. • Quicker but less accurate results. • Slower but better results. • Example: Farneback Algorithm • Example: Kanade-Lucas Algorithm 1 1 2 20 03 2 - KHUSHDEEP SINGH AND AAKASH VERMA 8

EXAMPLE OF SPARSE OPTICAL FLOW ESTIMATION 1 1 2 20 03 2 - KHUSHDEEP SINGH AND AAKASH VERMA 9

ITERATIVE KL OPTICAL FLOW COMPUTATION 1 1 2 20 03 2 - KHUSHDEEP SINGH AND AAKASH VERMA 10

E A N S R E OO N I U T T J C A A L I E E O F ET V G N D SI LG O R IT H M U 1 1 2 20 03 2 - KHUSHDEEP SINGH AND AAKASH VERMA 11

VOILA JONES ALGORITHM Efficient Visual Event Detection using Volumetric Features by Yan Ke, Rahul Suthankar, Martial Herbert [ICCV’ 05] 1 1 2 20 03 2 - KHUSHDEEP SINGH AND AAKASH VERMA 12

G N I T S O E V I T P A D A M E T A M A C H IN E L E O B A R N IN G A L G O R IT H M 1 1 2 20 03 2 - KHUSHDEEP SINGH AND AAKASH VERMA 13

1 1 02 2 03 2 Lecture on Ada. Boost by Jan Sochman, Jiri Matas KHUSHDEEP SINGH AND AAKASH VERMA 14

FOR THE GIVEN EXAMPLE 1 1 2 20 03 2 - KHUSHDEEP SINGH AND AAKASH VERMA 15

COMPARISON OF ADABOOST WITH OTHER METHODS Boos. Texter: A boosting-based system for text categorization by Robert E. Schapire and Yoram Singer. 1 1 2 20 03 2 - KHUSHDEEP SINGH AND AAKASH VERMA 16

ADABOOST PSEUDOCODE 1 1 2 20 03 2 - KHUSHDEEP SINGH AND AAKASH VERMA 17

SOME INITIAL RESULTS USING ONLY 1000 WEAK CLASSIFIERS No. of Training Examples No. of Features Misses False Alarms 200 2, 000 3% 60% 200 2, 000 19. 58% 30% 500 1, 000 12% 58% * TRAINED ON THE KTH DATASET 1 1 2 20 03 2 - KHUSHDEEP SINGH AND AAKASH VERMA 18