Online self-supervised learning for dynamic object segmentation. (April 2015)

• Record Type: Journal Article
• Authors: Guizilini, Vitor; Ramos, Fabio
• Is Part Of: International journal of robotics research. Volume 34:Number 4/5(2015)
• Page Start: 559
• Page End: 581
• Publication Date: 2015-04
• Subjects: Computer vision -- dynamic objects -- object segmentation -- machine learning -- Gaussian processes -- online learning; Robots -- Periodicals; Robots, Industrial -- Periodicals; 629.89205
• DOI: 10.1177/0278364914566514
• Languages: English
• ISSNs: 0278-3649

Abstract

This paper proposes a novel technique for the automatic segmentation of dynamic objects, solely using information from a single uncalibrated moving camera and without the need for manual labeling (or any human intervention, for that matter). Matching pairs of sparse features are extracted from subsequent frames, and the resulting optical flow information is divided into two classes (static or dynamic) using the RANSAC algorithm. This initial classification is then used to incrementally train a Gaussian process (GP) classifier that is then able to segment dynamic objects in new images. The GP hyperparameters are optimized online during navigation, with new data being gradually incorporated into the non-parametric model as it becomes available while redundant data is discarded, to maintain a near-constant computational cost. The result is a vector containing the probability that each pixel in the image belongs to a dynamic object, along with the corresponding uncertainty estimate of this classification. Experiments conducted using different robotic platforms, ranging from modified cars (driving at speeds of up to 50 km/h) to portable cameras (with a full six-degree-of-freedom range of motion), show promising results even in highly unstructured environments with cars, buses and pedestrians as dynamic objects. We also show how it is possible to cluster individual dynamic pixels into different object instances, and then further cluster those into semantically meaningful categories without any prior knowledge of the environment. Finally, we provide visual odometry results that testify to the proposed algorithm's ability to correctly segment (and then remove) dynamic objects from a scene, and how this translates into a more accurate motion estimate between frames.