Eigenspace-based Fall detection and activity recognition from motion templates and machine learning. We present a new spatio-temporal template: Motion Vector Flow Instance (MVFI)

1. Iván Gómez Conde

2. 1. Introduction
2. Related Work
3. Theory and Algorithms
 MILE Dataset
 Motion Vector Flow Instances (MVFI)
 Mathematics
4. Results
5. Conclusions
6. Further Research

3.  Automatically determining human actions and
gestures from videos or real-time cameras.
 New spation-temporal representation:
Motion Vector Flow Instance (MVFI)
 We compare it with other 2 motion templates:
silhoutte and Motion History Instance (MHI)
 Representations with a canonical transformation
with PCA and LDA.

4.  Automatically determining human actions and
gestures from videos or real-time cameras.
 New spation-temporal representation:
Motion Vector Flow Instance (MVFI)
 We compare it with other 2 motion templates:
silhoutte and Motion History Template (MHI)
 Representations with a canonical transformation
with PCA and LDA.

5.  Database of video scenes.
 6 human actions
 5 video sequences for each action
 12 different human subjects
 Sampling rate (25 frames/second)
 Static camera
 Videos were saved in AVI MPEG

6.  Walking
 Exaggerated Walking
 Jogging
 Bending over
 Lying Down
 Falling

7. Exagerated
Walking

8. Bending
over

9. Falling

10.  394 video sequences, 6 human actions, 12
different human subjects

11.  C++
 OpenCV (Open Source
Computer Vision)
 Python
 FFmpeg
 Octave
 Ubuntu

13. Training Testing

14.  M1: Silhouttes (foreground object)

15.  M2: Motion History Instance (MHI)

16.  M3: Motion Vector Flow Instances (MVFI)

17.  Video sequences with 3 types of motion
templates: m1, m2, m3

22. https://www.youtube.com/watch?v=sWY_mQ2_Gco

23.  Training set:
where is the image pertaining to the ith class and having
the jth frame within the sequence.

24.  The PCA canonical space is constructed from the
orthogonal vectors that possess the most variance
between all the pixels from the image sequence.
 Thus, with , we have:

25. This method projects the original images of the
sequences onto the new multidimensional space:

26.  We apply a Fisher criteria that maximizes the
between class variance and minimizes the within
class variance
 The Fisher linear discriminant function, , if
given by the ratio:

27.  We can write the corresponding eigenvalue
equation:
 The new orthogonal basis that takes the points
of the PCA space and transforms them to this new
space, we call the LDA space, through:

28. The points of PCA space are projected onto the new
multidimensional space:

29.  In this paper we performed an N-fold cross-
validation training process, where we constructed
all posible binary and multiclass combinations in
our dataset.
Training:
Testing:

30.  Two methods are used for determining the class of
an unknown test sequence:
 KNN clasifier
 SVM clasifier

31.  Summary of some representative training
statistics, showing the average number of
images, the average training times, and total run
times.

32.  Normalized histograms of recognition rates
obtained from the three different spatio-temporal
motion templates from

33.  Comparison of recognition rates for different
multiclass training: number of actions as a function
of different number of persons included in training.

34.  Comparison of recognition rates for different
motion templates as a function of incrementally
including more people in the training set.

35.  Comparison of motion templates for binary
classification of actions.

36.  The results of PCA and LDA training space with six
different action classes of this study.

37.  We performed a 10-fold cross validation for all six
action classes and all the training sequences of our
database.
 The Motion Vector Flow Instance outperforms all
other motion templates.

38.  This paper has compared two different motion
templates with a new spatio-temporal motion
template, MVFI.
 MVFI outperforms other methods for detecting
actions characterized by large velocities.
 This work suggest that it is important to preserve
velocity information in each image sequence.
 MVFI works well in all situations for action
recognition: different people, different clothing
types…
 Future studies shall consider both different camer
angles and distances.

41. http://www.youtube.com/watch?v=bNBMRNjIu_g&feature=player_embedded