Abstract:
The Army operates in complex, dynamic environments, without a priori information, at operational-tempos. Accurate state estimation requires an understanding of the dynami...Show MoreMetadata
Abstract:
The Army operates in complex, dynamic environments, without a priori information, at operational-tempos. Accurate state estimation requires an understanding of the dynamic elements in the scene. We propose a computationally efficient method for object classification in a dynamic scene suitable for a mobile platform. Certain phenomena of interest, such as trees waving or people walking produce repeatable and recognizable patterns when represented by an optic flow vector field. These patterns are invariant to a mobile observer's viewpoint. In this paper, we identify such patterns and demonstrate a computationally efficient framework to differentiate between dynamic and static background elements, mobile agents and ego-motion produced by the platform. This provides two significant benefits. The first being that when navigating through a dynamic environment, it is useful to have an understanding of the nature of the obstacles encountered, which may be used to determine which disturbances are benign and which ought to be avoided, without requiring burdensome analysis. The second is that by identifying which elements of the flow field are not due to egomotion, they can be removed from visual odometry calculations, allowing for robust state estimation in highly dynamic, army relevant environments. We offer experimental validation of the proposed methodology through a comparison of our classifier against ground truth and several other algorithms.
Date of Conference: 18-20 October 2016
Date Added to IEEE Xplore: 17 August 2017
ISBN Information:
Electronic ISSN: 2332-5615