Dance Movement Patterns Recognition (Part II)

Size: px

Start display at page:

Download "Dance Movement Patterns Recognition (Part II)"

Branden Hall
5 years ago
Views:

1 Dance Movement Patterns Recognition (Part II) Jesús Sánchez Morales

2 Contents Goals HMM Recognizing Simple Steps Recognizing Complex Patterns Auto Generation of Complex Patterns Graphs Test Bench Conclusions 2/32

3 Contents Goals HMM Recognizing Simple Steps Recognizing Complex Patterns Auto Generation of Complex Patterns Graphs Test Bench Conclusions 3/32

4 Goals Recognizing simple user s s movements Recognizing complex patterns Auto generation of reference patterns Pattern searching during the dance without any reference 4/32

5 Contents Goals HMM Recognizing Simple Steps Recognizing Complex Patterns Auto Generation of Complex Patterns Graphs Test Bench Conclusions 5/32

6 Hidden Markov Model (HMM) Highly used in Speech Recognition We adapted it for our problem Data sequences to analyze Stop move 1 move 2 stop HMM graphs for each recognition Viterbi s algorithm Initial State x1% x2% x3% State 1 (1-x2)% State 2 (1-x3)% Final State x5% (1-x1)% State 3 (1-x4)% (1-x5)% x4% 6/32

7 Contents Goals HMM Recognizing Simple Steps Recognizing Complex Patterns Auto Generation of Complex Patterns Graphs Test Bench Conclusions 7/32

8 Recognizing Simple Steps What we understand as being a simple step? Right, left, down, up, jump and twister How we divided these recognitions Horizontal Vertical Twister 8/32

9 Recognizing simple steps (Horizontal) Recognized steps Left step Right step Used data Horizonal variation of the centre of mass between frames User radius 9/32

10 Recognizing simple steps (Horizontal) Building a sequence and launching a thread Markov graph 10/32

11 Recognizing simple steps (Vertical) Recognized steps Up Down Jump Used data Vertical position of the centre of mass Vertical average position of the centre of mass User radius 11/32

12 Recognizing simple steps (Vertical) Building a sequence and launching a thread Markov graph 12/32

13 Recognizing simple steps (Twister) Recognized steps Twister Used data User radius Average radius Horizontal and vertical position of the centre of mass. 13/32

14 Recognizing simple steps (Twister) Building a sequence and launching a thread Markov graph 14/32

15 Contents Goals HMM Recognizing Simple Steps Recognizing Complex Patterns Auto Generation of Complex Patterns Graphs Test Bench Conclusions 15/32

16 Recognizing Complex Patterns What we understand as being complex pattern? Combinations of simple steps Left Step + Right Step + Left Step + Jump + Twister + + What we receive from the simple step recognition Step code Duration Step time Code: 11 Left Step Duration: 3 frames Time: second 54 16/32

17 Recognizing Complex Patterns Building a sequence and launching a thread Markov Graph. 17/32

18 Contents Goals HMM To Recognize Simple Steps To Recognize Complex Patterns Auto Generation of Complex Patterns Graphs Test Bench Conclusions 18/32

19 Auto Generation of Complex Patterns Graphs Easy way of building graphs for complex patterns recognition Included in the pattern recognition Saved as a text file 19/32

20 Auto Generation of Complex Patterns Graphs We receive the sequence of simple steps We build the graph for the recognition 20/32

21 Contents Goals HMM Recognizing Simple Steps Recognizing Complex Patterns Auto Generation of Complex Patterns Graphs Test Bench Conclusions 21/32

22 Test setup Test Bench Good external conditions Real time tests steps Possible results for each test Well detected Wrong detected Not detected 22/32

23 Test organization Test Bench Horizontal recognition Vertical recognition Twister recognition Complex pattern recognition 23/32

24 Test Bench (Horizontal recognition) High success rate No wrong detections In case of fast dance some steps get lost Individual steps Various steps (slow) Various steps (fast) 0,00% 0,00% 0,00% 0,00% 19,23% 0,00% 80,77% 100,00% 100,00% Well detected No detected Wrong detected 24/32

25 Test Bench (Vertical recognition) Very good individual step recognition Slow dance tests: some recognition problems Fast dance tests: recognition performance decreases Individual steps Various steps (slow) Various steps (fast) 12,50% 15,38% 18,18% 6,25% 0,00% 30,77% 81,82% 81,25% 53,85% Well detected No detected Wrong detected 25/32

26 Test Bench (Twister recognition) High recognition performance for steps separately taken Slow dance tests: start having major detection problems Fast dance tests: many steps are mixed Individual steps Various steps (slow) Various steps (fast) 16,67% 0,00% 16,67% 14,29% 16,67% 28,57% 83,33% 66,67% 57,14% Well detected No detected Wrong detected 26/32

27 Test Bench (Complex pattern) Perfect recognition performance for steps separately taken No concrete pattern to analyze Inherit problems Wrong simple step recognition Too slow simple step recognition Wrong received order 27/32

28 Contents Goals HMM Recognizing Simple Steps Recognizing Complex Patterns Auto Generation of Complex Patterns Graphs Test Bench Conclusions 28/32

radius variation due to arms movements Failed complex pattern

29 Conclusions (Difficulties encountered) Clothe variations Unknown frame rate Similarities in the vertical movements User radius variation due to arms movements Failed complex pattern recognition due to wrong order in simple movements recognition 29/32

30 Conclusions (Possible improvements) Visual detection Use of the data from the dance dance revolution pad Relating the beat detector with the user recognized steps Detection of other simple movements Development of a learning algorithm to improve the HMM graphs 30/32

31 Conclusions (Reached goals) We have found a good technique to recognize body movements In some cases the results have not been as good as we hoped but we think that can be improved This technique is also valid to detect more complex patterns We easily generate complex pattern graphs It has not been possible to search patterns without reference 31/32

32 Thank you very much! 32/32

Mobile Wireless Channel Dispersion State Model

Mobile Wireless Channel Dispersion State Model Enabling Cognitive Processing Situational Awareness Kenneth D. Brown Ph.D. Candidate EECS University of Kansas kenneth.brown@jhuapl.edu Dr. Glenn Prescott