Particle filtering is one of the key algorithms in the field of probabilistic robotics. They are specifcally useful in cases where the system's dynamic model and measurement functions are non-linear and non-Gaussian. In addition, they have been shown to perform well in cluttered scenes and in cases of short-duration occlusions. Particle filters are able to represent such distributions by representing a distribution by a set of "particles", which are a set of weighted samples.
The control system first needs to be fed with a target region to track. This can done by an instance segmentation algorithm, but for this study the target region is manually specifed by drawing a bounding box using mouse input. A stereo-depth camera attached to the user's AR glasses is used to capture an RGB-D image for each frame using an Intel RealSense D435i camera.
The following figure shows the output of the program at select frames. Video had a frame rate of 15 fps.
Program works well on irregularily shaped objects. Motion history can be derived from the predicted object tracks in the image frame.
