Real-Time Target Detection Method for UAV Embedded Platform

With the continuous development of UAV embedded platform in the direction of small size, low power consumption and low cost, the real-time object detection method deployed on it has become a research hotspot in the field of deep learning. Most of the existing mainstream target detection methods are based on large model design, which has the characteristics of complex architecture and huge calculation amount, and the detection accuracy is low when there is similar interference to the target in the UAV aerial image, which is difficult to deploy to the UAV embedded platform to meet the application requirements of real-time accurate detection. In this paper, a real-time object detection method deployed on embedded platform is proposed. Based on the idea of full convolutional architecture, a lightweight model based on twin network is designed, which makes the method have high real-time and detection accuracy. On the self-built ship data set, the detection accuracy can reach 87.35%, and the detection speed can reach 40FPS on the embedded platform, which meets the application requirements of real-time accurate detection.