Bi-heterogeneous Convolutional Neural Network for UAV-based dynamic scene classification
Jiewan Zheng, Xianbin Cao, Baochang Zhang, Yuanjun Huang, Yutao Hu · 2017
Unmanned Aerial Vehicle (UAV) plays a significant role in aeronautical surveillance. There often exists complex and dynamic natural scenes in surveillance videos, such as forest fire, ocean, landslide. Therefore, it is of great importance to classify dynamic scenes, which can facilitate object detection and tracking processes and improve the performance of visual surveillance. In this paper, a new Bi-heterogeneous convolutional neural network (Bi-CNN) method is proposed based on deep learning, which extracts both spatial and temporal information from a video to be exploited to decide which category the video belongs to. Considering the lack of proper datasets, we constructed a new dynamic scene dataset via UAVs. The experiments on several challenging datasets show that the proposed algorithm outperforms the state-of-the-art methods.