Research on object detection technology for human detection
Yanling Niu, Zhaohui Meng · Journal of Physics Conference Series · 2020
Abstract There are many classical application scenarios in computer vision, and human detection in dense scenarios is a classical problem. Its goal is to find all the people in an image or each frame of an video, as well as their positions and sizes, and frame them in a rectangular box. In recent years, deep learning has been applied to the field of computer vision, laying a solid foundation for human detection based on deep learning. Aiming at the crowdhuman dataset, based on faster-RCNN and feature pyramid network (FPN), this paper studies the training effect. In order to detect small objects, a multi-scale feature fusion network (MFFN) was proposed for feature extraction. According to the serious occlusion of dataset, a double-branch structure was proposed to improve the detection accuracy. These two structures combine to form a new network, we call it MDN, namely multi-feature fusion and double-branch network. Experiments show that these methods are helpful to improve the detection accuracy.