An anchor-based convolutional network for the near-surface camouflaged personnel detection of UAVs

With the lightweight and dexterous design, unmanned aerial vehicles (UAVs) are widely used to perform near-surface target detection missions in various complex environments. The camouflaged personnel detection in the images captured by the UAVs plays an essential role in the information acquisition, which determines the success of detection missions. However, the camouflaged personnel are not easily discovered due to the high similarity between the camouflage style and the background. In addition, there is only available few labeled sample images. To address above-mentioned problems, a camouflaged personnel dataset is first established, and a novel anchor-based method is then proposed to detect the camouflaged personnel. In addition, the efficient channel attention and the improved receptive fields block are added in the anchor-based convolutional network to focus on more features about the camouflaged targets. Besides, the non-maximum suppression is applied to determine the optimal bounding box on the targets. The quantitative results and visualization effects demonstrate that the mean average precision of the proposed method can reach 85%, and the recall can reach 83% on the developed dataset.