SCENE UNDERSTANDING: Toward a Safer Navigation. Damien VIVET ISAE-SUPAERO DEOS Toulouse France

Transcription

1 SCENE UNDERSTANDING: Toward a Safer Navigation Damien VIVET ISAE-SUPAERO DEOS Toulouse France

2 A three years old children is an expert of image analysis, content description and event recognition.

3 Our society is more technologically advanced than ever but still less advanced than a 3 years old child.

4

5

6

7 We already have excellent sensors

8 Why is it so hard?

9 One mission: To give computers visuals intelligence.

10 "Ingénieur ISAE-SUPAERO" (MsC) 15 Advanced Masters 6 Doctoral programs Alumnis / 1700 students 30% of international students 55 nationalities 1800 lecturers from the industry 200 professors and researchers 43 research directors (HDR)

11 To navigate we need three major capabilities : - perception - command law - localization (and mapping)

12 The visual perception: Man VS Machine

13 How to perceive the environment?

14 Humans also use landmarks to localize themselves

15 First localization and mapping based on landmarks

16 Modern cartographies

17 A landmark for a computer: keypoints

18 Applications of multi-view geometry Building Rome in a Day, S. Agarwal, N. Snavely, I. Simon, S. M. Seitz and R. Szeliski International Conference on Computer Vision, 2009, Kyoto, Japan.

19 Simultaneous Localization and Mapping: SLAM SLAM is a process in which a vehicle build a representation of its environment while localizing itself in this environment

20 Simultaneous Localization and Mapping: SLAM SLAM is a process in which a vehicle build a representation of its environment while localizing itself in this environment

21 Simultaneous Localization and Mapping: SLAM SLAM is a process in which a vehicle build a representation of its environment while localizing itself in this environment

22 Simultaneous Localization and Mapping: SLAM SLAM is a process in which a vehicle build a representation of its environment while localizing itself in this environment

23 Simultaneous Localization and Mapping: SLAM SLAM is a process in which a vehicle build a representation of its environment while localizing itself in this environment

24 Simultaneous Localization and Mapping: SLAM SLAM is a process in which a vehicle build a representation of its environment while localizing itself in this environment

25 Simultaneous Localization and Mapping: SLAM SLAM is a process in which a vehicle build a representation of its environment while localizing itself in this environment

26 Simultaneous Localization and Mapping: SLAM SLAM is a process in which a vehicle build a representation of its environment while localizing itself in this environment

27 Simultaneous Localization and Mapping: SLAM SLAM is a process in which a vehicle build a representation of its environment while localizing itself in this environment

28 Simultaneous Localization and Mapping: SLAM SLAM is a process in which a vehicle build a representation of its environment while localizing itself in this environment

29 Simultaneous Localization and Mapping: SLAM SLAM is a process in which a vehicle build a representation of its environment while localizing itself in this environment

30 Simultaneous Localization and Mapping: SLAM SLAM is a process in which a vehicle build a representation of its environment while localizing itself in this environment

31 Simultaneous Localization and Mapping: SLAM SLAM is a process in which a vehicle build a representation of its environment while localizing itself in this environment

32 Simultaneous Localization and Mapping: SLAM SLAM is a process in which a vehicle build a representation of its environment while localizing itself in this environment

33 Simultaneous Localization and Mapping: SLAM ISAE-SUPAERO, France: (application for car, bicycle and UAV)

34 Simultaneous Localization and Mapping: SLAM ISAE-SUPAERO, France

35 Mobile object detection and tracking: DATMO University of Coimbra, Portugal

36 Traffic monitoring LITIS (INSA Rouen), France and ISAE-SUPAERO Dr. Yadu Prabhakar (NAIT Edmonton)

37 "What to do next? How to go from standard geometry to semantics? We just have to be smarter...

38 "Road infrastructure detection

39 "Road infrastructure detection and mapping

40 "Road infrastructure detection and mapping

41 "Road infrastructure detection and mapping

42 "Road infrastructure detection and mapping

43 "Road infrastructure detection and mapping

44 "Road infrastructure detection and mapping

45 "Localization based on Road infrastructure detection

46 "Human behavior detection ISAE-SUPAERO and University of Orléans, France

47 "Abnormality detection ISAE-SUPAERO and University of Orléans, France

48 "How to go deeper?

49 "How to go deeper? A neural network is a simple assembly of connected elements called perceptron. The most you have, the Deeper you are.

50 "Road scene labelling University Paris-Est Chinese University of Hong Kong

51 Merci pour votre attention!!! Des questions? Damien Vivet