Spatial attention model based target detection for aerial robotic systems

Detecting interested targets on aerial robotic systems is a challenging task. Due to the long view distance of air-to-ground observation, the target size is small and the number is large in the scene. In addition, the target only occupies part of the image, and the complex background environment can easily cover the feature information of the target. In this paper, a novel target detection method based on spatial attention model is designed, which changes the existing methods to enhance the features of target areas by enhancing global semantic information. By learning the feature weights of different spatial locations in feature space, the method proposed can focus attention on the target regions of interest in an image, and suppress the background interference features, which enhances the feature information of the target regions, and deals with the class imbalance problem in detection. The experimental results show that the algorithm improves the detection accuracy of small air-to-ground targets and has a good detection effect for dense target areas. Compared with RefineDet, the state-of-art small target detector, our method can achieve better performance at a lower cost.