연속 프레임 간 변화 감지 기반 Feature Map 기억·재사용을 통한 ROI-CNN 추론 기법

초록

In autonomous driving systems, Convolutional Neural Networks are widely used for perception and decision making. Conventional inference executes heavy CNN computations on every frame without reusing results, causing inefficiency. ROI-CNN method reduces computation by processing only changed regions, but suffers from noise, boundary instability, and information loss. To overcome these limitations, we propose a memory-based ROI-CNN method that stores intermediate feature maps from previous frames in memory and reuses them while applying a filter to refine changed regions. This approach reduces redundant computations by selectively processing only dynamic regions, achieving a better trade-off between computational efficiency and prediction accuracy compared with conventional ROI-CNN. Experiments show that the proposed method reduces computational cost by about 29.66% compared with the conventional approach, while the RMSE increased only slightly by about 10.3%. This reduction in computation saves energy and improves the efficiency of autonomous vehicles. The method achieves a balanced trade-off between efficiency and accuracy essential for autonomous driving systems.

키워드