Active Semantic Distributed Perception

Abstract

We propose a novel framework for distributed perception that addresses the main challenges of today~s perception: observing semantic rather than metric entities, and being able to actively gather information about the environment. Imagine a scenario where humans and robots carrying heterogeneous sensors observe the world from very disparate viewpoints, and might be even at different times of the day or even the year. Observations like appearances in images or range measurements will vary dramatically in appearance posing extreme challenges not only to semantic detection (what is where) but most important in the data association cross agents. Traditional distributed systems have focused on the metric aspect (where is the robot or the target) with guaranteed detection and correspondence and Gaussian positional noise, and hence, cannot tackle the above described challenges. The fact that we have though multiple controllable agents allows us to: (i) apply active policies that will increase the information gain given semantic observations from heterogeneous sensors, and (ii) enforce a consistent data association and semantic consensus cross agents. Based on these two objectives of this proposal we will study and design algorithms for distributed active semantic perception.

Document Details

Document Type
DoD Grant Award
Publication Date
Jan 04, 2017
Source ID
N000141712093

Entities

People

  • Kostas Daniilidis

Organizations

  • Office of Naval Research
  • United States Navy
  • University of Pennsylvania

Tags

Fields of Study

  • Computer science

Readers

  • Computer Networking
  • Educational Psychology
  • Robotics and Automation.

Technology Areas

  • AI & ML
  • Autonomy
  • Autonomy - Autonomous System Control