Methods to copy human thinking

1. Deep learning for scene danger ranking

  • Overview

Motivation: Creating a machine with the skill of anticipating complex hazardous situations

Approach: DNN for hazard anticipation in every scene by embedding human experience

Function: Danger map for safe path planning with variable safety margins adjusted from level of danger

Applications: Autonomous driving, Urban assistant, Driving teacher

  • Description

  1. Golden standard of danger levels and location
    • Experiment designed by driving psychologists
    • Danger labels by driving instructors or by bio signals (ongoing) to avoid subjectivity. It enabled us to create a small dataset of scenes with level of hazard
  2. Train DNNs to perceive the world as drivers (object, contours, motion, prediction, saliency)
    • Incremental transfer learning of generic functions for reasoning about hazard in new scenes
    • Output dense map from few training pixels This approach enables hazard raking in every scene

  • Achievement

  • Feasibility to reproduce by DNN subjectively rated information: the first DNN of its kind
  • Multi-task fusion of transfer learning blocks was realized (FCN8s-Pyramid architecture) and analysis suggests that danger ranking benefits from transfer learning
  • Realizing the danger map with the function of prediction and anticipation of Hazards

2. Safe vehicle trajectory prediction using deep neural networks and camera images

3. Active learning from DNN mistakes