Target Detection and Classification via Multi-task Learning of Prediction and Classification

Detection and classification of target at early stage are critical in modern defense systems. Previous studies have primarily focused on binary classification between targets and non-targets, or single-task learning which is structurally limited in detecting and classifying targets simultaneously. This paper proposes a multi-task learning framework based on LSTM that jointly learns prediction and classification tasks, enabling both detection and classification within a single model. The model captures both the temporal patterns of target sequences and decision boundaries between target classes. To improve detection performance, we introduce a two-dimensional score vector that integrates prediction error and k-NN(k-nearest neighbor) distance, followed by Mahalanobis distance calculation. Experimental results show that the proposed anomaly score outperforms conventional methods on target detection. The model achieves high accuracy and macro F1-scores using full-length sequences and maintains reliable performance even with shorter input segments. These results confirm its potential for early stage classification during tracking.