Crowd Video Motion Capture by Concurrent Optimization of Shapes, Poses and Positions
Naoya Kajio, Atsushi Satito, Akihiro Sakurai, Ko Yamamoto · 2024
Pedestrian flow simulation is important to predict congestion in an urban area and prevent a crowd accident. Many studies have used the machine learning-based pedestrian flow model, which require a measurement of pedestrians to obtain training data. Not only the positional data of each person but also pose and body shape information is useful because it enables the model to learn pedestrian features implicitly including age, gender and social relationship. In this study, we present a video motion capture method that estimates correspondences of an unspecified number of pedestrians in different camera images using the body shape feature of Skinned Multi-Person Linear (SMPL) model. Simultaneously optimizing the correspondence, position, pose and body shape, we can find the same person in multiple cameras and reconstruct their poses and body shapes. We quantitatively compare the result of the method with that of an optical motion capture and qualitatively evaluate the method using an open dataset of pedestrians.