http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
임혜연(Hye-Youn Lim),안명수(MingShou An),강대성(Dae-Seong Kang) 한국정보기술학회 2024 한국정보기술학회논문지 Vol.22 No.3
Research on technology for detecting people in video has been steadily conducted, and various studies such as human tracking and behavior recognition for dense population analysis have recently been conducted. Pedestrian detection in dense crowds can cause problems such as poor accuracy and false detection due to cover or overlap between pedestrians. Therefore, this paper proposes an improvement method based on the YOLOv5 model to solve this problem. First, an attachment mechanism-based neural network was constructed in the network feature fusion stage to improve feature extraction performance and reduce the computational burden of non-maximum suppression(NMS). Next, we propose a structure to create and combine differential attention feature maps for overlapping targets by improving the resolution of pyramid features of each neural network layer through upsampling. Experiments on CrowdHuman datasets showed that the proposed method improved mAP 2.3% and loss 0.013 as a result of real-time detection tests.