Pixel Target Tracking
In 2012, we developed a novel algorithm for tracking pixel size targets. To track multiple targets simultaneously,
we maintain an extended Kalman filter (EKF) tracker for each target independently. To update each tracker with new observations, we
use nearest neighbor strategy, i.e. for each image, the point which is closest to prediction position is assigned as new observation. To
cope with missing pixels, we set a distance threshold. If the closest distance is larger than the threshold, we assume there is no new
observation. If a tracker does not have any new observations for a long time, its covariance matrix will be large and we drop/stop it.
Our system has the following key features:
- It can handle pixel targets.
- It can detect and track multiple targets simultaneously.
- It can use partial observations to track targets.
- It can track multiple groups of targets based on automatic camera coordination.
- It can actively adjust cam era view angle to continuously track targets when the targets fall out of the field of view.
- It can handle missing observations, say targets (on image) may disappear randomly.
- It can handle false alarms, say there are some image points which look like targets and used as observations.
- It can handle non-linear target trajectory.
- It can automatically initialize new target's 3D state purely based on camera observations.
- It can automatically rescue the tracking failure based on automatic initialization.
- It can handle observation noise.
- Plug-in scheme for camera sensors, i.e. it's free to add or remove any camera without changing to the system. Sensor appending or removing can be even done on the fly.
- It can automatically kill false targets based on confidence and kill redundant target trackers.
In short, our tracking system is fully automatic, robust, flexible, scalable, efficient and practical for multiple pixel targets tracking. With all these capabilities, we believe that our multiple camera based tracking system is able to work in real world situations and provides a better solution than radar based tracking system in some complex scenarios. The figure below shows an example of pixel target tracking.
Frame 500 with realistic rendering. Three axes are the targets (single pixel) and red balls are EKF trackers. There are multiple cameras tracking the targets.