Method for Real-Time Abnormal State Detection of Non-Cooperative Air Targets Based on Deep Learning

We propose a deep learning-based method for real-time abnormal state detection of non-cooperative air targets. Regarding the abnormal state detection task as a sequence labeling machine learning problem, the proposed method designs a deep temporal assessment network (DTAN). Firstly, DTAN integrates the target's trajectories and surveillance images as multimodal input time sequences, and employs MLP and ResNet separately to extract trajectory features and pose features at each time step. Secondly, a Transformer Encoder is incorporated to derive long-term target state features, followed by a CRF layer for temporal propagation constraints across time steps. Finally, the target's state (abnormal or normal) of each time step is predicted in real-time by the network. Besides, PolyLoss was employed to improve the training process of DTAN. Simulation experiments demonstrates that the proposed method surpasses conventional methods in accuracy, as well as reducing false alarm rate and decreasing assessment delay.