Power Line Detection by Pyramidal Patch Classification

Obstacle recognition, especially power line detection, is very important for the safety of unmanned aerial vehicle flight. Current methods for power line detection mainly rely on the assistance of spatial context, such as tower-line correlation. These methods tend to produce low detection rates without auxiliaries while high false alarm rates due to heavy clutters caused by complicated backgrounds. In this paper, we propose a pyramidal patch classification framework that explicitly excludes the clutters without any extra auxiliaries. This framework enables good balance between detection precision and time-critical requirement; thanks to the proposed hierarchical patches partition and selection strategy. Accordingly, we design a new spatial grid pooling layer for our convolutional-neural-networks-based classifier, which is trained on the set of pyramidal patches. The final power lines are obtained by line detection in each patch of the smallest size, coupled with a line-line correlation procedure. Our experiments show that the proposed method can eliminate most false alarms and obtain a high detection rate with low computational cost.