UAV Target Detection based on RetinaNet

The traditional target detection is basically impossible in the air-to-ground scene with high complexity. The detection algorithm based on deep learning has become a research hotspot with its applicability and robustness. However, the small scale of target and the lack of feature information in such scenes will make it difficult for UAV to effectively detect targets. Consider all these, this paper improves the network structure of feature extraction layer, re-selects the scale and quantity of anchors by system clustering method, and optimizes the calculation of Focal loss based on RetinaNet. Through simulation test, the method improves the detection accuracy under the premise of ensuring accuracy. Further more, several experiments on the UAV platform show that the improved RetinaNet has higher detection accuracy.