10 September 2024 in IEEE Sensors Journal
Most of deep learning-based maritime radar target detectors are trained in a full-supervised fashion, accurate annotations are required in the training stage. However, we may not label rightly some unforeseen targets for their unknown positions relative to the radar when the radar collects the target echoes for the detector training. The mislabeled targets in the training set would confuse the detector parameter optimization, decreasing significantly its performance. In this paper, we deal with the problem of detection with partial annotation, partial targets are mislabeled as the sea clutter in the setting. We propose a progressive learning (PL) framework, which guides correct supervised training of the detector by identifying the mislabeled target samples and relabeling them using the pseudo label. First, we pretrain the detector using the training set to separate the features of the clutters and the mislabeled targets based on objective echo differences between them. Then, we identify the mislabeled target samples in the high-dimensional feature space by the way of anomaly detection and clean them from the training set. Next, the detector is retrained utilizing the rest of the training set to obtain explicit decision boundary with intraclass compactness and interclass divergence through the sample-wise metric learning paradigm. Finally, we present a pseudo label assignment strategy to relabel the anomaly samples and optimize the detector parameters combining with the relabeled samples and the cleaned training set. Experimental results on public maritime radar database show the effectiveness of the proposed PL framework.
5 July 2024 in IEEE Transactions on Instrumentation and Measurement
This paper introduces a novel self-evolving architecture for radar target detection models, focusing on addressing generalization problem in intelligent target detection models operating within time-varying clutter environments. The architecture can continuously learn the time-varying clutter distribution features and dynamically evolve the radar detection models online with the aid of detection performance feedback, which can enhance the adaptive detection ability of radar system. Specifically, firstly, an intelligent detection model library is constructed by pre-training relevant models using offline samples gathered across diverse weather conditions. Subsequently, a three-level weighted scoring criterion is established to evaluate the performance of the radar detection models in real time and provide feedback information on the direction of radar self-evolution. Additionally, considering that the model retraining time is closely related to the sample size of the training set, a sample selection technique based on feature representation is proposed to select a small number of more representative samples to facilitate rapid incremental learning of the models. Finally, to validate the efficacy of the proposed approach, the performance of the radar’s online self-evolving detection is assessed using a publicly available dataset. The results of the evaluation confirm that the proposed architecture can improve the target detection performance by 9.04% within 1 minute in the face of time-varying clutter environment, which significantly solves the problem of poor detection performance caused by insufficient generalization of intelligent detection model.
15 October 2024 in IEEE Transactions on Aerospace and Electronic Systems
Classic maneuvering targets tracking (MTT) algorithms require accurate kinematic models, the performance of which would be degraded significantly for the intricate and abrupt maneuvers. In addition, their sensitivity to maneuverability would be reduced in the presence of clutters. In this paper, we consider the MTT problem without using preset kinematic models in the clutter environments, and propose a data-driven intelligent multi-frame joint tracking method (DIMJTM) for maneuvering targets with clutters. Initially, we introduce a maneuverability characterization method using modified Mahalanobis distance. Based on this, a backtracking vertex sampling strategy is proposed for selecting multi-frame historical track points matching the current maneuver pattern. Subsequently, we propose a pyramid shaped graph construction scheme to establish multiple spatio-temporal relationships among the selected track points and measurements of the current frame. Finally, the composed graph is fed into our designed interactive Transformer-graph attention network autoencoder (ITGAE). The simulated and real experimental results indicate that the DIMJTM not only accurately captures maneuverability changes, but also outperforms the mainstream IMM-PDA/IPDA algorithms and their improved variants in the single target tracking scenarios and the MHT/IMM-MHT algorithms in the multiple targets tracking scenario, achieving stable and high-precision tracking.
23 April 2025 in IEEE Transactions on Aerospace and Electronic Systems
Intelligent maritime radar target detectors, developed by deep learning technology, have shown encouraging detection performance. The detectors usually operate assuming that the training set and the test set are independent and identically distributed (i.i.d.). However, variable sea surface conditions can easily disrupt the assumption and significantly impair the detector's performance. This paper introduces the concept of unsupervised domain adaptation (UDA) to formulate the problem of detection in an unknown environment and proposes a UDA method, i.e., SL−UDA, based on the self-learning strategy to deal with it. The SL−UDA enables the detector to efficiently adapt to unknown environments while retaining previous knowledge gained from the existing data as much as possible. Specifically, the SL−UDA first trains the detector on the annotated data. Then, when the detector encounters an unknown environment, the SL-UDA generates the pseudo labels for current data by integrating adaptive confidence threshold and echo prior characteristics. Finally, the data with both label and pseudo label is used to retrain the detector, making the detector adapt to unknown environments while preserving previous knowledge. Experimental results on public maritime radar database show that the performance of the SL−UDA method is better than several classical UDA methods.
13 August 2025 in IEEE Transactions on Aerospace and Electronic Systems
Existing deep learning-based intelligent radar detection models generally suffer from insufficient self-evolution capabilities, leading to significant performance degradation in historical scenarios after training on updated scenarios. To address this challenge, this paper propose a two-stage training architectural framework based on knowledge distillation and fusion, designed to enable continuous self-evolution of radar detection models in dynamic scenarios through a dynamic alternation training between adaptation and refinement phases. Initially, an adaptive knowledge distillation algorithm is designed to unbiasedly distill the knowledge extracted by the model from historical scenarios, generating a compact core knowledge set and storing it in the knowledge buffer. Subsequently, in the adaptation stage, the model rapidly enhances its detection capability in updated scenarios by learning the feature distribution of the data noly from these updated scenarios, thereby acquiring additional detection knowledge. Finally, in the refinement stage, the model conducts the self-learning based feature alignment using the core knowledge set in the knowledge buffer, thereby restoring the detection performance in historical scenarios to achieve synergistic optimization between historical and updated scenarios. Additionally, by proposing the mutual information maximization loss to optimize the fusion process of historical and additional knowledge, the model is encouraged to learn robust detection features across scenarios. Experiments on public radar datasets demonstrate that the proposed method outperforms the advanced RDSEIL and GSS methods.
1 November 2026 in IEEE Transactions on Intelligent Transportation Systems