Enhancing Human Key Point Identification: A Comparative Study of the High-Resolution VICON Dataset and COCO Dataset Using BPNET

Abstract

Accurately identifying human key points is crucial for various applications, including activity recognition, pose estimation, and gait analysis. This study introduces a high-resolution dataset formed via the VICON motion capture system and three diverse 2D cameras. It facilitates the training of neural networks to estimate 2D key joint positions from images and videos. The study involved 25 healthy adults (17 males, 8 females), executing normal gait for 2 to 3 s. The VICON system captured 3D ground truth data, while the three 2D cameras collected images from different perspectives (0°, 45°, and 135°). The dataset was used to train the Body Pose Network (BPNET), a popular neural network model developed by NVIDIA TAO. Additionally, a comparison entails another BPNET model trained on the COCO 2017 dataset, featuring over 118,000 annotated images. Notably, the proposed dataset exhibited a higher level of accuracy (14.5%) than COCO 2017, despite comprising one-fourth of the image count (23,741 annotated image). This substantial reduction in data size translates to improvements in computational efficiency during model training. Furthermore, the unique dataset’s emphasis on gait and precise prediction of key joint positions during normal gait movements distinguish it from existing alternatives. This study has implications ranging from gait-based person identification, and non-invasive concussion detection through sports temporal analysis, to pathologic gait pattern identification.