Multi-View 3D Human Pose Tracking

This project implements the paper "Cross-View Tracking for Multi-Human 3D Pose Estimation at Over 100 FPS" (CVPR 2020).

Overview

This implementation focuses on multi-person 3D pose estimation and tracking from multiple camera views. The key features include:

• Cross-view tracking for robust 3D pose estimation
• Real-time performance optimization
• Support for multiple people tracking
• Handling of occlusions and view changes

Dependencies

• Python 3.7+
• PyTorch
• NumPy
• OpenCV
• Matplotlib
• tqdm

Dataset Download

You can download the Campus dataset from:

• Campus Dataset (OneDrive)

After downloading:

1. Extract the Campus_Seq1.tar.gz file
2. Place the extracted Campus_Seq1 folder in the project root directory
3. The dataset should follow the structure shown above

Note: The Campus dataset contains synchronized videos from three calibrated cameras, along with ground truth 3D pose annotations.

Dataset Structure

The code supports datasets in the following structure:

Implementation Details

The main components include:

• pose_matcher.py: Cross-view pose matching
• pose_estimator.py: 3D pose estimation from 2D detections
• display.py: Visualization utilities

Demo Results

Multi-View Camera Setup

The system utilizes multiple synchronized cameras to capture different viewpoints of the scene. This multi-view setup enables:

• Comprehensive coverage of the tracking area
• Better handling of occlusions
• More accurate 3D pose reconstruction

3D Pose Estimation Results

The visualization demonstrates:

• Accurate reconstruction of human poses in 3D space
• Real-time tracking of multiple subjects
• Robust pose estimation across different views

Key Features Demonstrated

1. Multi-view synchronized capture
2. Real-time 3D pose estimation
3. Robust person tracking
4. Occlusion handling
5. High-performance processing

References

This implementation is based on:

• Original Paper
• Official Dataset Repo
• Unofficial Implementation

Citation

[1] Chen, Long, et al. "Cross-View Tracking for Multi-Human 3D Pose Estimation at Over 100 FPS." The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.