Pedestrian tumble detection method based on lightweight human body posture estimation

• Main Authors: LI JIAHUI, SHI XIANGBIN, JUNG JUN-TAE, YAN ZHUO, YUAN LAOHU, FAN CHUNLONG, ZHANG DEYUAN, LIU BOYANG
• Format: Patent
• Language: Chinese; English
• Published: 16.08.2024
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

The invention provides a pedestrian tumble detection method based on lightweight human body posture estimation. The pedestrian tumble detection method comprises the following steps: acquiring a large number of pedestrian tumble pictures, and carrying out key point labeling and training set and verification set division; the method comprises the steps that a lightweight human body posture estimation model is constructed, trained and tested, the lightweight human body posture estimation model comprises a trunk, a neck and a head, the trunk is composed of a Repmix module, a ConvFFN module and a down-sampling module, the neck is composed of a DW-ViT module, and the head is composed of a convolutional layer; constructing a pedestrian tumble recognition model, and performing training and testing, wherein the pedestrian tumble recognition model is composed of two linear layers and two LAM modules; and performing pedestrian tumble detection by using the trained lightweight human body posture estimation model and the