Learning the state space based on flying pattern for bird detection

With the opening up of the low-altitude airspace, flying bird detection has recently attracted increasing attention in computer vision to avoid bird strike. However, compared with conventional object detection tasks, it is much more challenging to detect flying birds in aerial videos due to existence of small-sized targets, complexity of backgrounds with great variations and disturbances of bird-like objects. Motivated by the intuition of birds' flying periodicity, we propose a new method to solve the flying bird detection problem in a novel framework, termed flying pattern clustering (FPC), which learns key poses via AP-clustering algorithm, and then a Markov Model is established to describe the regular transition of key poses. As another advantage, a powerful Faster R-CNN model is used to obtain bounding box in each frame, which can significantly improve the flying bird detection performance. Experiments demonstrate that the FPC can achieve high detection accuracy and outperform state-of-the-art detection methods.