Multi-Pose Moving Human Detection based on Unmanned Aerial Vehicle in Real-Time

In this paper, a multi-pose human detection system is presented based on Unmanned Aerial Vehicle (UAV) imageries in real-time. A satisfying detection accuracy of multipose human detection in complex outdoor environments can be achieved with no accelerating hardware like graphics processing unit. First, 30,000 positive samples with various appearances are selected manually from different shooting angles and are used to train the convolutional neural network (CNN) classifier. Second, optical flow of the two successive frames taken by onboard camera is used to extract region of interests and a series of post-processing methods, i.e., average filtering, morphological operation and extraction of outer contour are applied to meet the real-time requirement. Then a two-stream CNN combining the appearance and motion information are applied to reduce the false detection accuracy. Field experiments in complex outdoor environments have been performed to verify that the proposed system can be used to detect multi-pose humans with high detection accuracy in real-time.