Loading presentation...

Present Remotely

Send the link below via email or IM


Present to your audience

Start remote presentation

  • Invited audience members will follow you as you navigate and present
  • People invited to a presentation do not need a Prezi account
  • This link expires 10 minutes after you close the presentation
  • A maximum of 30 users can follow your presentation
  • Learn more about this feature in our knowledge base article

Do you really want to delete this prezi?

Neither you, nor the coeditors you shared it with will be able to recover it again.



No description

Anupam Sobti

on 12 November 2016

Comments (0)

Please log in to add your comment.

Report abuse

Transcript of MAVI

SignBoard Detection
Restricted to White Text on Blue Signboard (as on campus)
Blue detection followed by White Detection inside the contour is used to confirm signboard presence.
Animal Detection
Safety concern for VI. Stray animals can be dangerous specially with a cane.
Cow Detection implemented using features from HoG.
Feasibility of using ground plane to improve accuracy in progress.
Dog detection in progress.
Texture Detection
Mobility Assistant for Visually Impaired
Face Detection and Recognition
Indispensable for social interaction between visually impaired individuals.
Implemented in Software (OpenCV).
Implemented on Zedboard (both HW/SW).
Location data is used to map signboards/hazards on the cloud or to extract information from Maps.
Mobile App
Currently receives a string using Bluetooth containing data from all the modules and displays it on the device.
Speaks out alerts like Pothole, Face Detection and names of recognized faces.
Server communication in progress for lookup of Signboard with coordinates.
Targets outdoor navigation for pedestrian users only.
Not a replacement for cane, but complements the cane.
Focuses on following aspects (based on user feedback)
Texture/ Surface detection
Animal detection
Signboard detection
Face detection and recognition

Control System
An initial implementation for both Ubuntu and Zedboard is available.
Being used to collate data from different modules and send to the mobile app.
Efforts being made to make the controller smarter to optimize the system for power/performance during runtime.
Pothole Detection
Major safety concern.
Using both the depth and RGB images as input, an anomaly in the ground plane is found and reported as a pothole.
Shadow Removal
Shadows cause inaccuracies in Texture Detection, Signboard Detection, Pothole Detection.
Current implementation works on deriving the illuminant invariant image but efforts are being made to develop an alternate method.
1. OCR for handwritten and Hindi literals.
Trying Deep Neural Nets on Embedded platform.
2. Shadows decrease detection accuracy.
Shadow Removal being taken up as an individual problem.
1. Execution time ~ 3.4 sec
- Hardware acceleration being done. Latest estimate (~1.7 sec)
2. Recognition accuracy is poor.
- 65% for upto 3m distance.
3. Shadows and position cause detection to be missed out
1. Position of animals creates a challenge for detection through image.
- Ground plane detection being explored for improvement in accuracy.
2. Relative position could be of help in declaring severity of danger.
1. Road vs Mud classification fails using SFTA features.
- Color features being included for classification
2. High execution time
- Gabor filter and other techniques being explored
3. Lower accuracy of pothole detection due to limited number of windows
- Using depth in conjunction
Intended to suggest walking directions
Also being used for pothole detection using discontinuities in texture.
SFTA = Segmentation Based Fractal Texture Analysis
1. Depth capture in daylight is erroneous due to IR Cameras and RGB is erroneous in the night
- Consistent images for the same frame are difficult to obtain
2. Training data is limited
1. Location accuracy is still limited to ~3m
2. Modules using localization results get impacted.
Left face not detected.
Only left face detected.
No face detected.
Features not coming out explicitly in illuminant invariant images.
Outdoor scenario with multiple colors
Images in indoor controlled environment
Images with Shadow
Illuminant Invariant Images
Images with Shadow
Illuminant Invariant Images
Front View : Face Detection
Side View: Body Detection
Full transcript