Pedestrian POSE estimation using multi-branched deep learning pose net.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Pedestrian pose and orientation estimation is crucial for human activity recognition and behavioral analysis.
• Accurately determining a pedestrian's focus versus their direction of travel presents a significant challenge in computer vision.
• Automated analysis of pedestrian behavior and intention requires robust pose and orientation estimation techniques.

Purpose of the Study:

• To propose and evaluate a deep-learning-based approach for accurate full-body pedestrian pose and orientation estimation.
• To introduce the multi-branched deep learning pose net (MBDLP-Net) for enhanced pose estimation and classification.
• To demonstrate the effectiveness of MBDLP-Net across multiple diverse datasets.

Main Methods:

• A deep-learning-based supervised model, MBDLP-Net, was developed for full-body pose and orientation estimation.
• The model was trained independently on the CIFAR-100 dataset.
• Performance was evaluated using three independent datasets: Body Orientation Dataset (BDBO), PKU-Reid, and TUD Multiview Pedestrians.

Main Results:

• The MBDLP-Net achieved a mean accuracy of 0.95% for full-body pose estimation on the BDBO and PKU-Reid datasets.
• On the TUD Multiview Pedestrians dataset, the proposed technique reached a mean accuracy of 0.97%.
• The results demonstrate the model's capability to efficiently distinguish full-body poses and orientations in various configurations.

Conclusions:

• The proposed MBDLP-Net effectively estimates pedestrian full-body pose and orientation.
• The approach shows superior performance compared to existing state-of-the-art methodologies.
• This technique offers a robust solution for automated pedestrian behavior and intention analysis.