Particle Filter for Closed-Loop Detection and Tracking from Occluded Airborne Images

This paper describes a novel vision-based target tracking and landing method that uses aerial images from an on-board camera. The proposed method explicitly deals with occlusions that often occur during these maneuvers. Normalized cross correlation (NCC) is used to locate an image patch in a reference image with a measure of certainty. The key insight is that over the course of the vehicle approach, there is a transition between the target being contained in the camera images, and camera images being contained in the target image. When a vehicle is at high altitude, the NCC of the target over an entire camera image is computed. When at low altitude, the reverse operation is performed: the NCC of the camera images is computed over the target image. Additionally, at both high and low altitude, we find interesting region using contour trees, and the NCC of the template with the region is calculated. This way, we can recognize a target even when it is only partly in view. A particle filter is used to fuse highly multi-modal measurements from the three techniques. Each particle chooses its update measurement using a roulette wheel selection with the size of the slice being proportional to the measurement's NCC and, therefore, converges to a location that has a greater NCC and numerous positive hits. The particle filter allows estimation of target position and velocity states, which are used to determine criteria for safe landing. We evaluate our system with an image-in-the-loop simulation and closed-loop flight tests with a quadrotor.