Binaural Sonification of Disparity Maps

1
Binaural Sonificationof Disparity Maps

• Alfonso Alba, Carlos Zubieta, Edgar Arce
• Facultad de Ciencias
• Universidad Autónoma de San Luis Potosí

2
Contents

• Project description
• Estimation of disparity maps
• Segmentation of disparity maps
• Object sonification
• Test application
• Preliminary results
• Future work

3
Project description

• The goal of this project is to develop a scene
  sonification system for the visually impaired.
• Images from a stereo camera pair will be used to
  detect objects in the scene and estimate the
  distance between them and the subject.
• A binaural audio signal will be synthesized for
  each object, so that the subject can hear the
  objects in the scene in their corresponding
  locations.

4
Scene sonification system

• The system will consist of the following stages
• Stereo image acquisition
• Disparity map estimation
• Disparity map segmentation (object detection)
• Binaural sonification of objects in the scene
• Here we will focus only on the segmentation of a
  given disparity map, and sonification stages.

5
Estimation of disparity maps

• Images from a pair of cameras, separated by a
  certain distance, form a stereo image pair.
• The position of a certain object in one of the
  images will be shifted in the other image by an
  amount inversely proportional to the distance
  between the object and the camera arrangement.
• This displacement is called disparity, and can be
  computed for each pixel to form a disparity map.
• We are currently working on a technique to
  compute disparity maps in realtime.

6
Segmentation of disparity maps

• Given a disparity map D(x,y), we perform a seeded
  region-growing segmentation to detect the objects
  in the scene.
• To choose the seeds, the algorithm uses a fitness
  measure given bywhere N(x,y) is the set of
  nearest-neighbors of (x,y), and q is a quality
  parameter (increases robustness to noise).
• This measure favors homogeneous regions (low
  dq)with the highest disparity (nearest objects).

7
Region-growing algorithm
1
Take a pixel from a regions border.
1
0
1
For each unlabeled neighbor, compare its
intensity to the regions average intensity.
1
0
1
If they are similar enough, include the neighbor
in the region.
1
0
1
8
Object sonification

• Sound coming from a specific location will suffer
  a series of degradations before it reaches our
  ears.
• These degradations provide various cues that our
  brain uses to locate the sound sorce.
• Binaural spatialization attempts to model these
  cues, in order to allow the listener to hear a
  sound as if it were coming from a specific point
  in space, which is typically defined in spherical
  coordinates (see below).

9
Object sonification

• We represent each object in the scene with a
  ping-like sound whose frequency depends on the
  disparity, so that the sound becomes more
  alerting as the object becomes closer.
• The audio signal corresponding to each object is
  fed through a binaural spatialization system
  whose parameters depend on the objects
  position.
• Spatialization is performed by modeling azimuth
  and range cues. Elevation cues have not been
  implemented (yet).

10
Azimuth cues

• Inter-aural Time Difference
• The sound source is delayed by a different amount
  for each ear
• Tn a a sin(q), Tf a aq.
• Inter-aural Level Difference (head-shadow)
• The sound is attenuated when passing through the
  head.
• This cue can be modeled with a one-pole one-zero
  filter

Brown et al., 1998
11
Range cues

• Artificial Reverberation
• Reverberation is the result of a large number of
  echoes originated from the reflection of the
  sound in flat surfaces such as walls.
• The level of reverberation is roughly constant
  and independent of source location.
• We use a simple model composed of 4 parallel
  delay lines with feedback.
• Attenuation
• The audio signal is attenuated according to the
  inverse quadratic law.
• The ratio between the signal and reverberation
  levels provides an additional cue for range.

12
Test application

• We simulate a moving scene by taking a 160 x 100
  sub-frame of a precomputed disparity map.
• The 10 most relevant objects are segmented but
  only objects that are near enough are sonified.

13
Preliminary results

• Fast segmentation times
• 5 ms per 160 x 100 frame in a 2.4 GHz dual core
  CPU
• Over 100 frames per second including sonification
  stage (but without disparity map estimation)
• Embedded implementation is viable
• Good azimuth representation object direction is
  easily perceived.
• Object range is perceived in a relative manner
  (e.g., one object is nearer than another), but
  not in an absolute way.
• Between 3 and 5 objects can be sonified before
  too much clutter is heard.

14
Future work

• Camera setup and calibration
• Realtime estimation of disparity maps
• Elevation cues in binaural spatialization
• Optimization of sonification system
• Implementation in an embedded device

15
Thank you!