wenbowen123/BundleTrack

What it solves

BundleTrack provides a way to track the 6D pose (position and orientation) of novel objects in real-time without requiring a 3D CAD model of the specific object or its general category. This removes the dependency on offline training or template matching based on pre-existing 3D models, which is a common limitation in most 6D pose tracking systems.

How it works

The framework combines three main components to maintain spatiotemporal consistency and reduce drift:

1. Deep Learning for Segmentation: It uses a video segmentation network to isolate the target object from the background.
2. Robust Feature Extraction: It leverages deep learning to extract features from the object.
3. Memory-Augmented Pose Graph Optimization: It optimizes the object's pose over time, using a memory system to ensure the tracking remains stable even during significant occlusions or fast object motions.

Who it’s for

This project is designed for robotics researchers and developers working on object tracking and perception, specifically those who need to track unknown objects in dynamic environments where 3D models are unavailable.

Highlights

• Model-Free Tracking: Does not require instance-level or category-level 3D CAD models.
• Real-Time Performance: Achieves a processing speed of 10Hz using a CUDA-based implementation.
• Robustness: Capable of long-term tracking with low drift, even under challenging scenarios like significant occlusions.
• Benchmark Proven: Outperforms state-of-the-art category-level 6D tracking and dynamic SLAM methods on public benchmarks.