An Attractive Motion Tracker
A novel magnetic motion tracking technique has been paired with deep learning to overcome some limitations of optical tracking systems.
Tracking motion is most commonly achieved with the help of an optical camera. However, systems that make use of cameras suffer from problems when the object to be tracked is either partially or fully occluded by other objects in the frame. These problems are particularly pronounced in applications such as finger tracking, where fingers are often hidden by nearby fingers, the hand, arm, or other objects.
In light of these issues, a research team has developed a new method, named IM3D+, that operates on the basis of deep learning and a new magnetic tracking principle. A flexible magnetic flux sensor array is used to measure magnetism from markers (inductor-capacitor (LC) coils) that are placed on the object to be tracked. A deep neural network was trained to translate these measurements into 3D representations of the markers in space. IM3D+ operates at 100 Hz to provide smooth motion reconstructions.
Magnetic tracking systems suffer from the so-called dead-angle problem, in which measurements cannot be taken if the marker is perpendicular to the driving coil’s flux direction. The team overcame this obstacle by implementing a structure-aware temporal bilateral filter that computes the weighting of time-series data based on sensor data. This filter is able to accurately reconstruct the missing data.
In addition to tracking finger movements, the researchers also see applications in tracking the flow of opaque fluids and small animals which may burrow — both scenarios in which optical camera-based solutions perform poorly.
Experiments were conducted to compare the accuracy of IM3D+ with conventional motion tracking approaches. A conventional method showed a mean square error of 26 square millimeters, whereas IM3D+ reduced that error to 14 square millimeters.
The new method overcomes some problematic elements of existing techniques, however those advancements come at a cost. IM3D+ requires significant setup, and specific placement of the flux sensors that are application-dependent. It also requires that markers be placed on the object under observation, which may be undesirable or impractical for some applications.
