Haojie Li Jinhui Tang Si Wu Yongdong Zhang Shouxun Lin Automatic Detection and Analysis of Player Action in Moving Background Sports Video Sequences IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 20, NO. 3, MARCH 2010
Introduction Global motion estimation Player body shape segmentation Analysis of action Experimental results OUTLINE
Presents a system for automatically detecting and analyzing complex player actions in moving background sports video sequences Providing kinematic measurements for coach assistance and performance improvement Video-based approach : low cost, no interference to the performance of players, can analyze the rich archived video clips INTRODUCTION
Block diagram of the proposed system INTRODUCTION Video Sequence Global motion estimation Action clips detection Action recognition Player shape segmentation Visual analysis Kinematic analysis Highlights library 1. The detected highlights are stored into library as video summaries for user’s quick browsing 2. Action recognize using CHMM ( continuous hidden markov models ) Action-based video indexing Kinematic parameters
GLOBAL MOTION ESTIMATION Global motion parameter
GLOBAL MOTION ESTIMATION
[24] N. Ostu, “A threshold selection method from gray level histogram,” IEEE Trans. Syst. Man. Cybern., vol. 9, no. 1, pp. 62–66, Jan
GLOBAL MOTION ESTIMATION Global motion vectors between two frames Outlier filtering Aligned image using estimated GME parameters Difference image between (b) (e) Background is accurately aligned
PLAYER BODY SHAPE SEGMENTATION … … Consecutive 2L-1 frames d1d2 ∩
Results of Algo1 PLAYER BODY SHAPE SEGMENTATION Problem: Work well only when object has apparent motion Reason: Doesn’t consider the object motion between frames
PLAYER BODY SHAPE SEGMENTATION Only selected key-frames are used to construct background … … Consecutive 2L-1 frames … … kf 1 kf L1 Kf -1 kf L2 L1L1 L2L2
Results of Algo2 PLAYER BODY SHAPE SEGMENTATION Th1 = 4 TH2 = 50 L = 11
Kinematic analysis : we present an automatic method through 2-D articulated human body model fitting, to get the joint angles. S=( x, y, θ, θ 1, θ 2, θ 3, d ) THE ANALYSIS OF ACTION Human body model Test body shape Edge map Distance transform map [35] The initial parameter is refine by searching with annealed particle algorithm[34] Global position & rotation parameter neck, hip, knee angle
Visual analysis Motion Panorama Overlay composition THE ANALYSIS OF ACTION Temporal median filtering Compare actions performed by different players or by the same player at different time No constraint that two clips should be of same scene
Global motion estimation EXPERIMENTAL RESULTS ( Interframe transformation fidelity ) Aligned image to I k by global motion compensated on I k-1 RANSAC & LTS with refinement procedures
Player body segmentation EXPERIMENTAL RESULTS
Action recognition EXPERIMENTAL RESULTS
Kinematic analysis EXPERIMENTAL RESULTS
Visual analysis EXPERIMENTAL RESULTS
Experiments on jump videos EXPERIMENTAL RESULTS