적응형 키프레임 문턱값 기반 영상-관성 오도메트리 (Adaptive Keyframe-Threshold Based Visual-Inertial Odometry)

서지정보

· 발행기관 : 제어·로봇·시스템학회
· 수록지 정보 : 제어.로봇.시스템학회 논문지 / 26권 / 9호 / 747 ~ 753페이지
· 저자명 : 김예준, 정재형, 박찬국

초록

In this paper, we apply a keyframe selection strategy that applies a 2-way marginalization method to a MSCKF(Multi-State Constraint Kalman Filter), which converts a sliding window to consist of keyframes. In SLAM(Simultaneous Localization and Mapping) or VIO(Visual Inertial Odometry) problems, various criteria for keyframes are used, such as relative translation, rotation, and tracking rate. However, each criterion has thresholds that determine whether the frame is a keyframe, and setting all of the thresholds is a cumbersome task. In this paper, the only criteria considered were those that are ratios of the number of newly observed features in that frame to all currently tracking features, since these criteria can substitute for other criteria. It was also found that the optimal threshold changed according to the mean feature tracking length, and the relation between them was determined via simulations and fitted with a polynomial. Adaptive keyframe threshold selection using the fitted polynomial was then applied to the trajectory whose mean feature tracking length varied in the simulation and KITTI dataset. The results indicate that the position error was decreased compared to the fixed keyframe threshold method.

영어초록

In this paper, we apply a keyframe selection strategy that applies a 2-way marginalization method to a MSCKF(Multi-State Constraint Kalman Filter), which converts a sliding window to consist of keyframes. In SLAM(Simultaneous Localization and Mapping) or VIO(Visual Inertial Odometry) problems, various criteria for keyframes are used, such as relative translation, rotation, and tracking rate. However, each criterion has thresholds that determine whether the frame is a keyframe, and setting all of the thresholds is a cumbersome task. In this paper, the only criteria considered were those that are ratios of the number of newly observed features in that frame to all currently tracking features, since these criteria can substitute for other criteria. It was also found that the optimal threshold changed according to the mean feature tracking length, and the relation between them was determined via simulations and fitted with a polynomial. Adaptive keyframe threshold selection using the fitted polynomial was then applied to the trajectory whose mean feature tracking length varied in the simulation and KITTI dataset. The results indicate that the position error was decreased compared to the fixed keyframe threshold method.

참고자료

· 없음