关键词: 无人机, 深度学习, 目标检测, 特征融合, 感受野

Abstract: Aiming at the problems of large changes in the scale of aerial images of unmanned aerial vehicles (UAV), great recognition difficulties, and generally small targets, the paper proposes an object detection algorithm for UAV aerial photography based on improved SSD RCBnet. In order to improve the feature extraction capability of the network, the algorithm modifies the feature extraction network of the SSD algorithm to Resnet 50 and adopts the feature fusion method to fuse the feature maps, and uses the fused feature maps to build a feature pyramid. In addition, in order to enhance the algorithm's ability of objects detecting, a multi scale convolution structure of attention mechanism is designed to effectively adjust the receptive field and realize the parallel operation of the feature map of the convolution kernel of different sizes. Aiming at the problem of extremely imbalanced positive and negative samples in training, the algorithm uses the Focal Loss function to train the network model so as to focus on difficult samples. Compared with other classic algorithms, the algorithm proposed in the paper has higher detection accuracy, better detection performance and robustness in UAV aerial images. Compared with SSD, the accuracy is improved by 3.46%.

Key words: unmanned aerial vehicles, deep learning, object detection, feature fusion, receptive fields