We propose a successive convex matching method to detect actions in videos' The proposed scheme does

1
Successive Convex Matching for Action
Detection Hao Jiang, Mark S. Drew and Ze-Nian Li
School of Computing Science, Simon Fraser
University, Vancouver BC, Canada V5A 1S6
We propose a successive convex matching
method to detect actions in videos. The proposed
scheme does not need foreground/background
separation, works in strong clutter and cases in
which objects have large deformation. The
proposed scheme converts the hard non-convex
problem into a sequence of much easier convex
programs.

Introduction
Property 1. If matching cost
surfaces are convex, the LP exactly solves the
continuous extension of the discrete matching
problem. Property 2. For general problems, the
LP approximates the original cost surfaces with
their lower convex hulls. Property 3. We
need only several basis target points
(corresponding to the lower convex hull vertices)
to construct the LP. Property 4. Using Simplex
Method there are at most 3 non-zero target points
for each site.
Fig. 9. Finding actions in a standard dataset.
(a, b, c, d) Templates and sample results.
To improve the approximation, we use the
following successive convexification scheme
Table 1. Detection Confusion Matrix
Delaunay Triangulation of feature points on
template images
Set initial trust region for each site the same
size as target image
Calculate matching costs for all possible
candidate target points
Find lower convex hull vertices in trust
regions and target point basis sets
Update trust regions
Action Detection
Build and solve LP relaxation
Update control points
Trust region small?
No
Template
Output results
Yes
Fig. 5. Diagram of successive convexification.
Fig. 10. Finding gestures in sign language
sequence.
Fig. 11. Finding actions in an indoor sequence.
Fig. 1. Detecting actions in videos.
Comparison of SC-LP, ICM and BP with ground
truth data
Experimental Results
Actions can be represented as a sequence of
postures with temporal constraints. Action
detection is carried out by posture sequence
matching and template-target similarity
comparison. Matching a template posture sequence
to video is formulated as an optimization
problem The above problem is
usually highly non-convex. If d(.) uses an L1
norm, we can relax it into linear programming
(LP)
Method
Fig. 12. Finding actions in a baseball sequence.
Fig. 3 Cost surface, lower convex hull and basis
labels.
Fig. 6. Random sequence matching.
Conclusion

• Propose a successive convex programming scheme to
  match video sequences using intra-frame and
  inter-frame constrained local features.
• The proposed scheme involves a very small number
  of basis points and thus can be applied to
  problems that involve a large number of target
  points.
• It has been successfully applied to locating
  specific actions in video sequences.

Fig. 2. Template to video sequence matching.
Fig. 7. Random sequence matching with duplicated
objects.
Matching in clutter
Fig. 8. (c, d) Target image features and edge
maps (e, f) SC-LP matching result (g, h)
Greedy scheme result (I, j) Chamfer matching
result (k, l) BP result.
For further information Please contact hjiangb,
mark, li_at_cs.sfu.edu.
Fig. 4 Searching process of LP.