Efficient Hierarchical Markov Random Fields for Object Detection on a Mobile Robot

TL;DR

This paper introduces two hierarchical Markov Random Field models for real-time object detection on mobile robots, improving multi-object distinction and achieving up to 11 fps on standard hardware.

Contribution

The paper presents novel hierarchical MRF models and efficient optimization techniques for fast, multi-class object detection on mobile robots.

Findings

Achieved near-realtime performance of 11 fps.

Effectively distinguished overlapping objects.

Validated on robot obstacle course footage.

Abstract

Object detection and classification using video is necessary for intelligent planning and navigation on a mobile robot. However, current methods can be too slow or not sufficient for distinguishing multiple classes. Techniques that rely on binary (foreground/background) labels incorrectly identify areas with multiple overlapping objects as single segment. We propose two Hierarchical Markov Random Field models in efforts to distinguish connected objects using tiered, binary label sets. Near-realtime performance has been achieved using efficient optimization methods which runs up to 11 frames per second on a dual core 2.2 Ghz processor. Evaluation of both models is done using footage taken from a robot obstacle course at the 2010 Intelligent Ground Vehicle Competition.