High-density pedestrian detection algorithm based on deep information fusion

• Published in: Applied intelligence (Dordrecht, Netherlands) Vol. 52; no. 13; pp. 15483 - 15495
• Main Authors: Zhang, Hexiang, Yang, Xiaofang, Hu, Ziyu, Hao, Ruoxin, Gao, Zehang, Wang, Jianhao
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.10.2022; Springer Nature B.V
• Subjects: Algorithms; Artificial Intelligence; Clustering; Computer Science; Data analysis; Data integration; Density; Iterative methods; Machines; Manufacturing; Mechanical Engineering; Pedestrians; Processes; Target detection; Deep information fusion; Pedestrian detection; High density; Clustering algorithm
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-022-03354-1

In order to improve the accuracy of high-density population detection, a high density pedestrian detection algorithm (YOLOv4-HDPD) is proposed based on deep information fusion. By increasing the connection points of cross-layer fusion, high-level semantic information is further integrated with feature information. The improved Iterative Self-Organizing Data Analysis algorithm (ISODATA) makes the anchor value more suitable for the network model without increasing the number of parameters. Moreover, the network anti-interference ability is increased by replacing the CIOU algorithm target detection object. Compared with the original network, the YOLOv4-HDPD network has improved in m A P and a v g I O U . Under the premise that the detection speed of the network is basically not affected, m A P is increased by 5.28% and a v g I O U is increased by 5.73%. In terms of the current results, the network algorithm has been improved the detection effect of high-density pedestrians. At the same time, the network provides a new idea for solving the clustering and detection of dense targets in real scenes.