Dynamic Semantic Motion Segmentation

Spring 2016, Undergraduate Researcher, IIIT Hyderabad

Worked on implementing incremental surface reconstruction and tracking in dynamic scenes.

  • To segment moving objects in a given frame, we used epipolar constraint with stereo camera and joint inference with semantics using Dense CRF formulation for improving motion labelling.
  • For tracking and fusion, we used InfiniTAM library. The library was used to return rotational (R) and translational (T) parameters between two consecutive frames from KITTI car dataset.
  • We used these parameters to represent the scene in a voxel grid. The idea was to fuse and find the surface of the moving object in an incremental fashion in these voxels after every frame using TSDF ray casting.
  • We used keypoint matching using SIFT features, and R & T parameters from the library to reconstruct the static scene by taking a weighted sum of both.
  • We applied Medioni’s transformation to reconstruct the object in motion, separately.

The project could not achieve the desired results as the R & T so obtained in voxel space were only accurate up to a scale, given less features on a moving object and hence, point cloud registration failed for moving objects.