MULTICLOD: Multiclass object detection

Computer vision for smart cars and safer roads

Demonstrations

Ladder-DenseNets for semantic segmentation

  • This demonstration presents semantic segmentation results obtained by Ladder-DenseNet-121
  • the presented accuracy (74.3 mIoU) can be obtained for around 133 ms (7.5 Hz) on Titan X
  • we have presented the demonstration at:
    • Computer vision for Road Scene Understanding workshop ICCV CVRSUAD 2017
    • Croatian Computer Vision Conference CCVW 2017
    • conference AI2Future 2017
Ladder DenseNet

Inferring the traversability map by semantic segmentation

  • We apply the inverse perspective transform to the semantic map recovered by Ladder-DenseNet-121 (74.3 mIoU, 7.5Hz on Titan X).
  • the resulting bird-eye semantic view (top-left subfigure in both figures on the right) allows to detect drivable parts of the environment
  • the technique could support the following reasoning within an autonomous driving system
    • the vehicle can navigate over purple regions in the map (road)
    • exceptionally, the vehicle can prudently drive over pink and light green regions in the map (sidewalk, terrain)
    • all other colours represent various kinds of obstacles (pedestrians, cars, etc.)

traversability map traversability map

Detection of road-safety attributes in video

  • We have designed a convolutional architecture for detection of events in continuous video and applied it to detect road-safety attributes
  • we focus on the merge-lane attribute:
    • the positive section is annotated with a red line in the satellite view (right-top)
    • for illustration, we show a negative image (right-middle) and a positive image (right-bottom)
  • we have trained the model on around 5000 positive and 5000 negative sequences of 25 images:
    • positive sequences are extracted at around 150 physical locations with merge lanes (manual annotation was performed by a third party)
    • we obtained negative sequences by random sampling, at least 50m away from all annotated merge-lane locations;
  • we achieved 94% AP on independent test-set (check the video)
  • post-hoc analysis shows that most false negatives have been caused by mis-annotationed data
  • conclusion: the proposed approach can be used to correct human annotation errors

Google maps negative example positive example

Closed-loop control of a robotic car

  • We have developed closed loop controllers for the robotic car Loox which was purchased in the scope of the project VISTA.
  • Our controllers run in the spearate thread and directly modulate the current going to the two driving motors and the motor which operates the steering wheel.
  • The desired parameters (speed, steering angle) can be configured throughout a high-level API.
  • We mapped API calls to keypresses so we are able to drive as in a computer game (check the video).

Loox Loox