Developing Neural Scene Understanding for Autonomous Robotic Missions in Realistic Environments
Abstract
Achieving rapid and robust understanding of scenes in complex and changing environments is critically important for autonomous robotic systems performing tasks to support realistic outdoor missions. Outdoor missions are scenario-oriented and are not just dependent on searches for objects. Autonomous missions need to detect time-sensitive scenarios that are signified by salient objects and environments providing mutual context in an evolving scene. Real-world, time-sensitive mission scenarios may include human activity and evolving threats, such as forest fires, chemical hazards, or adversary encroachment. We describe the implementation of a novel composite neural software configuration for complex outdoor missions that is tested against simulated test mission scenarios using two convolutional neural network (CNN) models separately trained on component objects and places image databases. Our proof-of-principle testing of the composite CNN supports the benefits of such a system when tuned to detecting time-sensitive scenarios that are keyed to the success of the mission. We examine five real-world mission scenarios to analyze the benefits of adding environmental data assimilation and physical modeling to improve neural scene understanding. We find that achieving autonomous inference of mission intent from detected and surveilled activities would make autonomous missions even more valuable.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 2018
- Accession Number
- AD1060448
Entities
People
- Arnold D. Tunick
- Ronald E. Meyers
Organizations
- United States Army Research Laboratory