탄두 파편데이터 검출시스템의 중복검출 제거를 위한 convLSTM 기반 필터 연구

Abstract

Warhead arena fragmentation test is a test used for analyzing a warhead’s lethal effect on the target by using fragmentation data such as speed, mass, and distribution. Detecting objects from a high-speed video that captures fragments just to penetrate a target plate in arena test, vision-based fragmentation detection system acquires fragment velocities and fragmentation distribution. Because only the first penetration data must be obtained to determine the velocity, in this study, the bounding box filter-based detector was used to remove the consecutive duplicate detection result. However, the bounding box filter has an additional dummy area surrounding the fragment area; if new detection occurs in this dummy area, it gets removed by the bounding box filter. This decreases the data acquisition rate and hinders a tester from obtaining sufficient fragmentation data for analyzing the warhead effect. In this paper, a pixel-level filter was proposed based on convolutional long short-term memory (convLSTM); the proposed filter utilizes a spatial association loss function for filtering out duplicate results. Owing to the spatial association loss function, which yields a strong correlation between successive frames, the convLSTM-based filter can generate pixel-level filter masks and remove duplicate results effectively. The experimental results demonstrated the superior performance of the filter.