(I) Objective: Develop a system to identify individuals climbing walls using pose estimation and action recognition techniques.

(II) Approach: Combined YOLOv7-Pose model, ByteTrack for object tracking, and an action recognition model for comprehensive analysis.

(III) Dataset: Utilized publicly available datasets, including COCO, CrowdPose, MPII, MHP, and AI Challenger 2017. 

• (Table.1) illustrates the various key-point annotations present in the public datasets. We utilized the 17 key-points defined by the COCO dataset to standardize and align other datasets for training Yolov7-Pose.

(Table.1) Yolov7-Pose training datasets

(Fig.1): The upper part of Figure1 shows the key-point performance of Yolov7-Pose on the COCO validation set, while the lower part displays the performance of ViTPose++. 

(Fig.2): The upper part of Figure2 shows the key-point performance of Yolov7-Pose on the OCHuman validation set, while the lower part displays the performance of ViTPose++. 

(Fig.3): The upper part of Figure3 shows the key-point performance of Yolov7-Pose on the MPII validation set, while the lower part displays the performance of ViTPose++. 

(Fig.4): The upper part of Figure4 shows the key-point performance of Yolov7-Pose on the AIC validation set, while the lower part displays the performance of ViTPose++.