고속 목표물 추적 시스템에서 Huber 기반 EKF 가속화를 위한 GPU 기반 선택적 병렬화 기법
GPU-Based Selective Parallelization for Accelerating Huber-Based EKFs in High-Speed Target Tracking Systems
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This paper deals with a Huber-based extended Kalman filter which can reduce the impact of outliers and enables more robust state estimation, but the additional computation causes real-time processing issues. To solve this problem, a GPU-based selective parallelization technique was proposed, and the goal is to maximize computational performance by optimizing the roles of CPU and GPU. In the proposed technique, the State Propagation and Huber Cost Function steps, which have high computational overhead, are performed on the CPU, and the Update Iteration and Initial Prediction steps, which have effective parallel operations, are designed to perform matrix multiplication and Gauss-Jordan matrix inverse calculation techniques in parallel. As a result of the simulation, the computation time was reduced by about 18.62% compared to CPU execution alone, and the speed was improved by 21.73% and 19.74% in Update Iteration and Initial Prediction, respectively. It was confirmed that the selective parallelization technique improves real-time performance.
Abstract
This paper deals with a Huber-based extended Kalman filter which can reduce the impact of outliers and enables more robust state estimation, but the additional computation causes real-time processing issues. To solve this problem, a GPU-based selective parallelization technique was proposed, and the goal is to maximize computational performance by optimizing the roles of CPU and GPU. In the proposed technique, the State Propagation and Huber Cost Function steps, which have high computational overhead, are performed on the CPU, and the Update Iteration and Initial Prediction steps, which have effective parallel operations, are designed to perform matrix multiplication and Gauss-Jordan matrix inverse calculation techniques in parallel. As a result of the simulation, the computation time was reduced by about 18.62% compared to CPU execution alone, and the speed was improved by 21.73% and 19.74% in Update Iteration and Initial Prediction, respectively. It was confirmed that the selective parallelization technique improves real-time performance.