Make a Mark in Motion Capture Without Markers

Varied applications ranging from virtual and augmented reality to physical rehabilitation and filmmaking require a detailed understanding of the movements of the human body. Whether this knowledge is needed to train a machine learning model, map physical movements into a virtual world, or otherwise, effective and practical data collection mechanisms are needed to capture it.

Presently, the most accurate motion capture systems make use of markers that are worn at key locations on the body, and a specialized camera system that tracks those markers. But wearing special suits covered in markers is cumbersome, and the tracking systems are expensive and complex. Furthermore, they can only capture data within a limited area of measurement, rendering them impractical for many use cases.

In an effort to move the state of the art forward, many research groups are currently developing so-called markerless motion capture systems. These systems may use cameras in conjunction with computer vision algorithms to detect the locations of the body’s joints in three-dimensional space, or they may incorporate accelerometers or other sensors to assist with capturing the data.

Whatever the technical details of a particular motion capture system may be, one question will always be at the forefront of the minds of its developers: just how good is it? Perhaps the best way to answer that question is to compare the new method against a gold standard system of today. Given the complexity and cost of these systems, that is not always practical.

But going forward, it will be much easier. A team at the University of Bath has just released a new dataset called BioCV. This dataset consists of video data paired with measurements from traditional motion capture systems. By having this treasure trove of data available at their fingertips, researchers can quickly compare the performance of their method to the best existing systems. And in addition to evaluating a new approach, the BioCV dataset also allows developers to rapidly iterate on new designs to refine their methods.

The BioCV dataset features synchronized data from nine HD cameras operating at 200 Hz, 3D marker trajectories from optical motion capture systems, force plate data, and photogrammetry scans for 15 participants performing controlled movements like walking, running, jumping, and hopping. Primary trials include synchronized marker and video data, while secondary trials focus solely on video-based motion capture. Compared to existing datasets like HumanEva and Human3.6M, BioCV uniquely includes force plate and photogrammetry data to aid in biomechanical analyses and body model creation. As such, this dataset supports the prediction of forces and joint moments, which is normally impractical outside of laboratory conditions.

For those interested in working with the BioCV dataset, it is provided free of charge. However, one must apply for access and agree to the researchers’ terms before they will hand the dataset over.