All research

Multimodal Terrain Classification for Off-Road Environments

Robust terrain understanding for the wild using depth-assisted semantic segmentation.

Jul 2021 Computer Vision

A depth-assisted, multimodal model for classifying unstructured off-road terrain.

  • Built a robust terrain-classification model using depth-assisted semantic segmentation.
  • Collected and generated a novel RGB-D + semantic-segmentation dataset of fully off-road / wild terrains.
  • Used SOTA models (Intel's Dense Prediction Transformer and MiDaS) for dataset benchmarking.
  • Designed a dual-stream RGB + Depth multimodal architecture with fusion-based feature combination.