Global data association for MOT using network flows

Posted on Jul 10, 2019

Define data association as a MAP problem. The problem is then mapped into a cost-flow network, and solved with a min-cost flow algorithm.

Basis of the mapping: there is an analogy between finding non-overlapping object trajectories and finding edge-disjoint paths in a graph.

MAP under non-overlap constraints

• $\x_i=(x_i,s_i,a_i,t_i)$: detection response
• $\calX=\{\x_i\}$: a set of object observations
• $T_k=\{\x_{k_1},\ldots,\x_{k_{l_k}}\}$: a single trajectory hypothesis
• $\calT=\{T_k\}$: an association hypothesis, a set of single trajectory hypotheses.

(differ time ? time increasing ??)

The objective of data association is to maximize the posterior probability of $\calT$ given the observation set $\calX$:

Reduce the search space by using the observation that one object can only belong to one trajectory:

Assume the motion of each object is independent,

A Bernoulli distribution is used to model the cases of an observation being a true detection as well as being a false alarm ($\beta_i$ is the probability for $\x_i$ being a false alarm):

$P(\cT_k)$ is modeled as a Markov chain, which includes initialization probability $P_{entr}$, termination probability $P_{exit}$, and transition probabilities $P_{link}(\x_{k_i+1}\mid \x_{k_i})$:

(still not global?)

Min-cost flow solution   This formulation can be mapped into a cost-flow network $G(\cX)$ with source $s$ and sink $t$.

Given an observation set $\cX$:

for every observation $\x_i\in \cX$,

• create two nodes $u_i,v_i$,
• create an arc $(u_i, v_i)$ with cost $c(u_i,v_i)=C_i$ and flow $f(u_i,v_i)=f_i$
• an arc $(s, u_i)$ with cost $c(s,u_i)=C_{en,i}$ and flow $f(s, u_i)=f_{en,i}$
• an arc $(v_i, t)$ with cost $c(v_i,t)=C_{ex,i}$ and flow $f(v_i, t)=f_{ex,i}$

for every transition $P_{link}(\x_j\mid \x_i)\neq 0$,

• create an arc $(v_i, u_j)$ with cost $c(v_i, u_j)=C_{i,j}$ and flow $f(v_i,u_j)=f_{i,j}$. • each flow path can be interpreted as an object trajectory
• the amount of the flow sent from $s$ to $t$ is equal to the number of object trajectories
• the total cost of the flow on $G$ corresponds to the (negative) log-likelihood of the association hypothesis
• the flow conservation constraint guarantees that no flow paths share a common edge

The optimal cost should be calculated over all possible $f(G)$, which is the amount of flow sent from source to sink. Published in categories Note