논문윤리하기 논문투고규정
  • 오늘 가입자수 0
  • 오늘 방문자수 135
  • 어제 방문자수 590
  • 총 방문자수 2790
2024-11-21 02:53am
논문지
HOME 자료실 > 논문지

발간년도 : [2023]

 
논문정보
논문명(한글) [Vol.18, No.1] Post-Processing Method of Deep-Learning Network to Improve People Counting Performance
논문투고자 Oh-Sung Kwon
논문내용 Accurate occupant information in a building is of paramount importance for prompt evacuation in the event of a fire. Accordingly, the establishment of a monitoring system using a thermal imaging sensor is preferred. In this case, the object must be distinguished and analyzed only by the volume of the thermal image area. In this paper, we propose analysis neural network YoloV5 and a post-processing method to supplement its recognition results. In general, neural networks often show overfitting or underfitting depending on the amount and nature of training samples. This phenomenon is due to the dependence of training samples and structural limitations of neural networks. Accordingly, the inference results of neural networks also often lead to unexpected results. In this paper, as a way to improve this, we propose an algorithm that corrects the misrecognized part through an image feature flow-based analysis and an overall system configuration plan. The proposed method analyzes the time series of images and quantifies the differences to correct the volume analysis error using the backtracking method. As a result of the experiment, the time delay due to the additional post-processing was not large, and it was confirmed that the recognition rate and performance improved by 21.28% compared to the existing method.
첨부논문
   18-1-09.pdf (973.5K) [1] DATE : 2023-03-03 09:11:01