| Title |
Moving Object Detection Using Multi-resolution Attention Mechanism |
| Authors |
하태길(TaeGil Ha) ; 임종인(JongIn Lim) ; 양동원(DongWon Yang) ; 이준희(JunHee Lee) ; 김도경(DoKyoung Kim) ; 최진영(JinYoung Choi) |
| DOI |
https://doi.org/10.5573/ieie.2019.56.3.81 |
| Keywords |
Background subtraction in a moving camera ; Moving object detection ; Multi-resolution ; Attention model |
| Abstract |
In this paper, we propose an effective image processing method to quickly detect small moving objects in a high-resolution image where cameras are not fixed. To detect small objects at a distance, high-resolution images should be used. However, in general, the size of the input image and the speed of the algorithm are inversely proportional to each other, and so the real-time speed can not be achieved in the high-resolution image. In order to achieve real-time speed, we propose a multi-resolution attention mechanism. The proposed method effectively processes high-resolution images using complexity-dependent attention mask for each input frame, where complex parts are processed with high resolution and simple parts are handled with low resolution via down-sampling. In addition, a multi-resolution motion compensation method is developed so that the method can be applied to a nonstationary camera. Experimental results show that the algorithm using the proposed multi-resolution attention mechanism achieves a speed improvement more than or equal to 85% and a speed of 6.67 fps compared to the existing state-of-the-art algorithm, which is the baseline of the proposed method, without significant degradation of performance in Full HD video. |