Multi-Cue Particle Filter Tracker

SIR+SIS Particle Filter Tracking
Mixed sampling mechanism (SIR+SIS) can extend the variety of candidates to deal with high-dimensional sparse peak searching.







Multi-Cue Likelihood Evaluation

Demo Video: 4D Object Tracking
(x, y, r, θ)

Candidate group behavior

Blue contours: SIS (from current color projection)

Green contours: SIR (from last candidate set)

Red contours: Best candidate as tracking result

Interaction System

²Finite State Machine: 4 states

³State Intent Detection

®Use Optical flow to detect whether user is rapidly moving target to trigger interaction

³State Object Learning

®Show “Checking Box” to let user put target in for information extracting

³State Motion Tracking

®Use Particle Filter to robustly track target and send signal to interaction program

³State Losing Handle

®If target is losing, wait for user putting back then continue interacting

Demo Video:
Interactive Game – Brick Breaker

Interaction system: 4 states switching

4 dimensional estimate: location x & y, size, rotation angle

Interactive Game Example

Summary

²Particle Filter tracking

³Advantage: High flexibilities of

®State dimension

®Motion model

®Adding new features & tune weights in tracker for different application scenarios

³Trade-off:

®Feature description vs. Computation time

®Number of particles ～ Complexity

²An entire visual interaction system is purposed and implemented for game applications