Pose Estimation for Pedestrain Upper Bodies in Crowds
نوع المنشور
ورقة مؤتمر
المؤلفون

This poster presents two approaches for improving human pose estimation from video data. The first approach enhances the BRC methodology by leveraging a dataset composed of three distinct feature types. Using a hold-out validation method with a 70/30 train-test split based on pedestrian IDs, the model achieves a Mean Squared Error (MSE) of 97.14. It generalizes well for head trajectory prediction but encounters challenges with fast shoulder detection, demonstrating improved performance on unseen data. The second approach adapts YOLOv8-pose for top-view human pose estimation by modifying it to work with overhead video data and reducing the standard 17 key points to only two per shoulder. A train-validation-test split of 70/15/15 is implemented, ensuring training and testing datasets originate from different videos within the same experiments. This approach achieves an MSE of 81.45, excelling in fast shoulder detection but requiring a larger dataset for optimal generalization. While it performs well in its designated task, its effectiveness slightly diminishes when applied to entirely new data. The comparative analysis highlights the strengths and limitations of both models in different aspects of human pose estimation, providing insights into their applicability for various real-world scenarios.

المؤتمر
عنوان المؤتمر
Traffic and granular flow 2024
دولة المؤتمر
فرنسا
تاريخ المؤتمر
2 ديسمبر، 2024 - 5 ديسمبر، 2024
راعي المؤتمر
Traffic and Granular Flow (TGF), Lyon, France