Particle Filter Tracking Architecture for use Onboard Unmanned Aerial Vehicles

Because of their ability to reach unique vantage points, rotary-winged unmanned aerial vehicles are well suited for visual target tracking missions. However, these missions are very burdensome for the vehicle operator since he must interpret the incoming video frames, update the orientation of the camera, and update the position of the vehicle to ensure the target is successfully tracked. This research seeks to provide an automated system architecture, which will perform these three tasks with minimal operator input. The majority of the research is focused on the video interpretation task. A particle filter is chosen to estimate the position of the target within each video frame. The particle filter is a recursive, sample-based state estimation tool that is capable of approximating non-Gaussian distributions governed by nonlinear models. Much of the recent particle filter research can be classified into three categories, data fusion, efficiency improvement, and color model adaptation. The data fusion works describe methods to blend information from various sensors in a structured manner. Most of the research relies upon static systems, where the fusion and other measurement parameters remain constant. However, the few exceptions show the promise of adaptive fusion systems, where the parameters change based on the tracking conditions. The efficiency improvement works describe methods to handle the particle filter’s inherently large computational load to achieve real-time performance. It is typically handled by changing the number of particles based upon properties of the particle distribution. The color model adaptation works allow the reference color model to change as the tracking conditions change. This allows more accurate color measurments to be made. The proposed system introduces a real-time, adaptive particle filter, which fuses information from multiple sensors. The parameters of the particle filter, including the number of particles, are all controlled using properties of the particle distribution. The output of the