Multiple Target Tracking Using Spatio-Temporal Monte Carlo Markov Chain Data Association Qian Yu, Gerard Medioni, and Isaac Cohen Edwin Lei

Previous Work Assumes one-to-one mapping between targets and observations. This is not accurate because – Observations correspond to blobs, not points – Occlusion results in a single object being detected as multiple objects

General Idea For a region of interest, find the best cover based on spatio-temporal consistency in both motion and appearance

Formulation Let Y be the set of all available foreground regions within [1,T], where T Є N + Let a track be a path traveled by a target Let the set of all covers be ω=(τ 0, τ 1,…, τ k ), where τ 0 is the set of false alarms and τ k is a rectangular cover for the kth track

Formulation The problem is to infer an unknown number of k tracks given the set of observations Y Mathematically, we want to solve Hard to solve

Monte Carlo Markov Chain Data Association Approach Construct a Markov chain which will converge to the target distribution p(ω|Y) Data-driven proposal of new covers Two types of proposed covers (moves) – Temporal: form consistent tracks – Spatial: change the cover at one instant

MCMC Data Association Symmetric temporal information forwards and backwards in time -> bi-directional sampling With p(ω|Y), solving becomes easy

Results