Flight Maneuver Recognition Through Modified YOLO Neural Network

Abstract

Flight Maneuver Recognition (FMR) involves automatically recognizing patterns in air target maneuvers. Precise and real-time recognition of air target maneuvers from flight data is crucial for promptly analyzing and responding to the combat strategies employed by enemy air targets. Due to uncontrollable factors affecting the duration of flight maneuvers, the sampled data from the flight state parameters varies in length. Neural networks are extensively utilized in FMR due to the capacity of RNN's hidden states and CNNld's convolution kernel to handle input data with varying lengths. However, existing neural network models for FMR primarily concentrate on the classification task while neglecting the segmentation task related to flight data. This paper presents a modified YOLO-based multi-task learning neural network for FMR that enables neural networks to learn segmentation and classification tasks concurrently. Furthermore, an LSTM neural network was trained using the same simulated datasets, and comparative experiments show comparable performance in classifying flight data sampling points for both neural networks.