Coarse-to-fine object detection in unmanned aerial vehicle imagery using lightweight convolutional neural network and deep motion saliency

Abstract

Unmanned aerial vehicles (UAVs) have been widely applied to various fields, facing mass imagery data, object detection in UAV imagery is under extensive research for its significant status in both theoretical study and practical applications. In order to achieve the accurate object detection in UAV imagery on the premise of real-time processing, a coarse-to-fine object detection method for UAV imagery using lightweight convolutional neural network (CNN) and deep motion saliency is proposed in this paper. The proposed method includes three steps: (1) Key frame extraction using image similarity measurement is performed on the UAV imagery to accelerate the successive object detection procedure; (2) Deep features are extracted by PeleeNet, a lightweight CNN, to achieve the coarse object detection on the key frames; (3) LiteFlowNet and objects prior knowledge is utilized to analyze the deep motion saliency map, which further helps to refine the detection results. The detection results on key frames propagate to the temporally nearest non-key frames to achieve the fine detection. Five experiments are conducted to verify the effectiveness of the proposed method on Stanford drone dataset (SDD). The experimental results demonstrate that the proposed method can achieve comparable detection speed but superior accuracy to six state-of-the-art methods.