A Real-Time Detection Drone Algorithm Based on Instance Semantic Segmentation

With the rapid development of drones, drones are widely used in various fields and bring convenience to people's production and life. However, they also bring security problems to society and the country. Especially in airports or military areas, the flight of drones can cause some problems. In order to effectively supervise the drone, this paper proposes a real-time detection drone algorithm HR-YOLACT which is based on instance semantic segmentation, and designed a new drone data set. The proposed algorithm combines the real-time instance semantic segmentation algorithm YOLACT with the deep high-resolution representation classification network HRNet. Firstly, feature maps are extracted by HRNet's backbone network. Secondly, the feature pyramid network is used to further extract image features, so that the network has better classification ability. Finally, the improved prediction head is utilized to detect the boxes of drones. In addition, this paper uses cross entropy instead of focal loss as the loss function to obtain better network training speed and quality. The experimental results show that HR-YOLACT has faster detection speed and higher detection precision than existing popular real-time object detection and real-time instance semantic segmentation algorithms.