Current projectile technology is at a point where most systems are either unguided, or guided via GPS to stationary targets. In order to engage moving targets and/or navigate in GPS-denied environments, infrared or electro-optical seeker technology must be employed. Our goal is to autonomously detect and track moving or stationary targets onboard a small, gun-launched projectile using a strapdown electro-optical seeker without the use of GPS. In this paper we propose two object detection algorithms that can be rapidly trained, and implemented onboard representative hardware to accomplish this goal. The first detection method is based on randomized Ferns, a non-hierarchical classification scheme which generates a discriminative classifier suitable for in-flight adaptation. We compare this to a baseline template matcher using normalized cross-correlation. We also describe how the projectile state estimation algorithms necessary for closed loop guidance can be coupled with targeting algorithms to provide rotation and scale estimates, decrease search window area to speed up the detection rate, and reduce the probability of false detections. We demonstrate the effectiveness of this approach by performing a series of experiments in which we rapidly train a detector and then generate real-time target measurements for line-of-sight rate estimation on embedded hardware.
展开▼
