Upload
buck-joseph
View
217
Download
0
Tags:
Embed Size (px)
Citation preview
Everybody needs somebody: Modeling social and grouping behavior on a linear
programming multiple people trackerLaura Leal-Taix´e, Gerard Pons-Moll and
Bodo RosenhahnICCV2011
Outline
• Goal• Multiple people tracking• Modeling social behavior• Experimental results• Conclusion
Goal
• People detection is not always correct.• It is important to merge the detection results
into right trajectoies.
Multiple people tracking
• divided in two steps– object detection– data associationform complete trajectories
• Build a graph with the nodes pedestrian detections
• The matching problem is equivalent to minimum-cost network flow problem
Multiple people tracking
• ,trajectory of k• Find the that best explains the
detection.• 4
• P(oi|T) is the likelihood.
Multiple people tracking
• trajectory Tk have following dependencies– Constant velocity assumption find oi depends on oi-1,oi-2
– Grouping behavior – Avoidance term
Multiple people tracking
• Three kinds of edges:– Link edges– Detection edges– Entrance and exit edges
Multiple people tracking
• Link edges• The edges (ei, bj) connect the end nodes ei
with the beginning nodes bj in following frames,with cost Ci,j and flag fi,j
• Flag =1 if oi and oj belong to Tk,and ∆f≤Fmax
• 111
Multiple people tracking
• Detection edges• The edges (bi, ei) connect the beginning node
bi and end node ei, with cost Ci and flag fi
Modeling social behavior
• If a pedestrian doesn’t meet any obstacles, he will naturally follow a straight line.
• But the pedestrian will have some social behavior.
• Add Social Force Model (SFM)and Group behavior(GR) into the problem.
Modeling social behavior
• Social forces have three main terms:– The desire to maintain certain speed– The desire to keep away from others– The desire to reach a destination
• We focus on first two!
Modeling social behavior
• Constant velocity assumpion– When a person walk at a speed V at time t– We assume he will have speed V at time t+∆t
Modeling social behavior
• From the training sequence in [22] , we learn the probabilty of Pg and Pi
[22] S. Pellegrini, A. Ess, K. Schindler, and L. van Gool. You’ll never walk alone: modeling social behavior for multi-target tracking. ICCV, 2009. 1, 2, 5, 7
Experimental results
• To show the importance of social behavior and the robustness of our algorithm at low frame rates, we track at 2.5fps (taking one every tenth frame).
Experimental results
• DA (detection accuracy)• TA (tracking accuracy)• DP (detection precision)• TP (tracking precision)