Automatic Target Recognition XXXVI

We present a physics-based framework for automatic target recognition that combines the robustness of the Generalized Likelihood Ratio Test (GLRT) with a new sequential estimation strategy. GLRT is a proven detection method that also provides parameter estimates, but its high computational cost limits real-time use. Our approach introduces “superparameters”—intermediate variables capturing key aspects of the target state—that are estimated first, reducing complexity before full parameter recovery. This enables low-latency, low-power detection and characterization of targets in EO/IR and other sensing modalities, while maintaining GLRT’s performance in low-SNR conditions. The method is applicable to defense and surveillance missions requiring both accurate detection and rapid parameter estimation in operational environments.

Digital image enhancement is a field that incorporates prediction and estimation based on local and global features and the entire image. Artificial intelligent techniques require training based on known information and can be a good tool for image enhancement, however, if the learning process is incomplete or corrupted based on faulty and irrelevant images these techniques can faulter. In this work we investigate a robust predictive technique, the Kalman filter, KF, often used for tracking and guidance applications to evaluate its performance for image enhancement and recovery. To expand our work, we have taken a few liberties with the KF and utilized a Bayesian approach along with a Hidden Markov Model estimator to improve the performance of the KF. Other statistical techniques have also been investigated based on probability density function estimation and its application with the KF to improve the filter’s performance given that the image’s structural information may be utilized for these purposes.

This paper examines the effects of denoising on radar target identification. Three different denoising techniques are examined 1) Denoising using wavelet decomposition, 2) Denoising using Convolutional Autoencoders, and 3) Denoising using generative adversarial networks. The paper also addresses the question whether any denoising improves target recognition performance. Real stepped-frequency radar is used to test these denoising algorithms. The data represents backscatter from commercial aircraft models as recorded in a compact range. The issue of azimuth ambiguity is also considered. Computational cost and speed are also considered. Time needed to denoise the radar backscatter is of particular importance because of the urgency to make an identification decision in a real battle scenario.

Deep neural networks are powerful tools that allow one to achieve state of the art performance for target recognition on SAR. Recent literature on SAR target recognition has heavily pursued the use of large, complex neural networks trained on increasingly larger classification datasets. To justify the use of these models, a common assumption is that SAR datasets are sufficiently difficult to learn. Here, we explore this basis. By using classifiers like K-Nearest Neighbor in conjunction with simple image processing, we first show that basic ML models can achieve similar accuracies to DNNs. These accuracies give a baseline for evaluating how much complex models help for improving accuracy on SAR target recognition. Overall, our findings motivate new questions around the cost benefit tradeoffs of large models and the true difficulty of the common benchmark datasets.

Compromised target recognition systems could potentially leak important information about their training data. Preventing such a compromise is critical in maintaining operational security. As a potential mitigation, differential privacy methods can provide theoretical guarantees on dataset privacy. In this work, we examine how differentially private methodologies can be integrated into an ATR pipeline. In particular, we focus on the process of generating ATR templates that are differentially private as a means of protecting hypothetically sensitive training data. We provide an analysis of synthetic aperture radar templates using a distributional model, common in statistical ATR methods such as multinomial pattern matching. Our work provides an analysis of both the template generation itself as well as statistical tests derived from the generated templates, and we observe the steep costs of high-dimensional templates. By understanding the effect differential privacy has on template generation, we can begin to characterize the performance/privacy trade-offs of differentially private ATR.

Understanding the relationships between data points in the latent decision space derived by the deep learning system is critical to evaluating and interpreting the performance of the system on real world data. Detecting “Out- of-Distribution” (OOD) data for deep learning systems continues to be an active research topic. We investigate nonparametric online and batch approaches for estimating distributional separation or “outlierness”. Using open source simulated and measured Synthetic Aperture RADAR (SAR) datasets, we empirically demonstrate that the concepts of OOD and “Out-of-Task” are not synonymous.

Out-of-distribution (OOD) detection is an important part of automatic target recognition (ATR) systems. The capability to reject unknown classes improves reliability and trust in an ATR, and permits the use of otherwise closed-set classifiers where open-set recognition is necessary. In this paper we present multinomial feature matching (MFM), a method for detecting OOD data in the latent feature space of neural classifiers, and apply it to models trained on the SAMPLE+ dataset. We show that MFM has efficient and low-overhead runtime characteristics, and that it exhibits good performance when applied to OOD targets in SAMPLE+ and other SAR datasets, including both vehicle targets and clutter.

Automatic Target Recognition (ATR) in tactical and remote sensing scenarios frequently encounters novel target classes absent from labeled training data, while multimodal sensors (EO, SAR, IR, radar) offer complementary cues for robust detection. Conventional unimodal or supervised methods struggle to discover and categorize these unseen targets in unlabeled streams. We propose a cross-modal novel class discovery framework leveraging a JEPA-based approach to learn independent embeddings across multiple modalities, enabling label-free identification of novel targets. Modality-specific encoders generate parallel embeddings and project them into latent space. Novel classes are discovered through sustained energy spikes in unimodal temporal prediction and spikes in cross-modal prediction. The framework provides novel class discovery resilient to sensor noise, occlusion, and deception. This work advances scalable, adaptive ATR for dynamic environments with emerging threats.

Target detection (TD) using Radar Range-Doppler Maps (RDMs) is a longstanding and critical challenge within the defense sector. Classical target detection approaches can struggle in cluttered environments. With the recent advancements of Vision-Language Models (VLMs), we explore an out-of-domain application of VLMs for target detection that leverages contextual information from human operators in the field. To enable this, we developed a scalable synthetic data generation pipeline for training and evaluation across two tasks: pointing and counting. We illustrate its effectiveness through numerical experiments by comparing an off-the-shelf VLM with a fine-tuned version. Our results show promise of this approach for enhancing radar-based target detection.

Understanding Earth’s dynamic changes requires automated recognition systems capable of temporal reasoning over remote sensing data. Extending the IARPA SMART heavy-construction dataset, we integrate newly generated text captions to describe evolving geospatial events. By fusing geo-INT image sequences, we train a multimodal large language model (MLLM) to generate descriptive captions and predict future target states. Our unique GPT-based transformer architecture includes a 7-billion-parameter model fine-tuned from a general foundation. We address challenges of sparse satellite collections through data-efficient training and visual question answering (VQA) data augmentation. Building upon prior work in SAR-based ATR and VQA, our results demonstrate that modern MLLMs can generalize across time-dependent sensing domains, advancing geospatial-temporal analysis for improved safety, security, and sustainability.

The authors investigate detection, classification, indexing, and automatic tracking of multiple stationary and moving objects in complex dynamic environments. We conduct fundamental and applied research, as well as undertaking technology developments in machine learning and pattern recognition aimed for intelligence, surveillance, and reconnaissance (ISR). On mini unmanned aerial systems, the proof-of-concept onboard actionable intelligence-centric machine vision module is implemented using ultra-compact low-power single-board computer, cameras, and transceiver. The developed learning algorithms are processed by single-stage end-to-end fully convolutional neural networks (CNNs). The modified variants of CNNs are investigated. The novelty lies in the use of new informative and regularizable loss functions with consequent hyperparameter optimization, as well as devised postprocessing observers aimed to perform robust indexing and dynamic tracking. The data-driven algorithms and factorizations yield trustworthy learning models, improving overall capabilities and enable decision superiority. For quantitative assessment, the ISR-quantifiable metrics are found in relevant environments.

Operational synthetic aperture radar (SAR) automatic target recognition (ATR) requires rejecting non-target objects while maintaining classification accuracy on known targets, but standard models trained on the MSTAR and SAMPLE datasets are prone to learning background shortcuts that undermine both goals. We establish this through target–background decomposition and dual saliency analysis: baseline domain classifiers achieve high accuracy by reading background texture, not target signatures. To correct this, we develop an integrated multi-task ATR pipeline combining chimera augmentation, base triplet loss, and chimera paired triplet mining, validated through a component ablation. The multi-task ATR pipeline redirects the domain head from background to target, and critically, domain accuracy survives even as domain clustering vanishes in the embedding space, indicating the domain head has shifted from background texture to target-level scattering differences. Baseline models also achieve perfect false alarm detection, but via background texture. A stress test breaks this ceiling, confirming that the pipeline’s detection is grounded in target signatures rather than background cues.

Long-range perception at high frame rates remains a challenge for autonomous systems, with limited open-source resources. We present a long-range, narrow-FOV flash LiDAR dataset for 2D/3D vehicle detection/segmentation: over 70,000 frames at 20 Hz up to 1000 m, covering eight vehicle classes (ground vehicles and a light aircraft). Frames include 3D point clouds, fused range–intensity images, five per-point features, ground-truth 2D bounding boxes, 3D cuboids, and semantic masks. We release a reproducible benchmarking suite with preprocessing, training, and evaluation scripts in an open-source repository. Included are standard 2D/3D detection and segmentation DNNs trained and tested on the dataset. We evaluate detection and classification accuracy across object classes and range bins, along with throughput and latency for real-time use. Results demonstrate robust long-range perception using the LiDAR system and reveal sensitivities to range, target aspect, point density, and SNR. We also include pipelines for vehicle tracking, multi-view registration for 360° modeling, and embedded deployment. Dataset and code will be publicly released to support advances in long-range ATR.

Detecting fast-moving objects in dynamic scenes is a fundamental challenge in event-based vision, with applications ranging from terrestrial surveillance to space-based tracking of Resident Space Objects (RSOs). Conventional frame-based imaging and convolutional pipelines struggle to meet the latency and power requirements of resource-constrained platforms. In this work, we present a scalable method for generating labeled training data by superimposing real event streams from stationary-camera foreground recordings and ego-motion background recordings, and demonstrate that a lightweight CNN trained on this data generalizes to real, unmodified event streams. This provides a practical foundation for future deployment in autonomous detection systems, including space-based applications where event cameras are well-suited to the constraints of spacecraft payloads.

Event-based cameras (EBCs) encode pixel-level radiance changes with microsecond latency and high dynamic range, enabling robust motion extraction while suppressing static background clutter. However, their sparse, asynchronous output requires algorithms that operate directly in event space. We present Frequency Rate Information for Event-Space (FRIES), a neuromorphic framework that detects periodic structure in event streams to discriminate man-made objects, such as drones, via rotor and vibration signatures. FRIES temporally filters events, forms activity-based regions of interest, and applies localized Fourier analysis to extract dominant frequencies, which are then tracked using a Resonant Time Surface. Indoor and outdoor experiments demonstrate recovery of mechanical chopper and drone rotor rates from ~74–451 Hz, highlighting frequency-domain event processing as a promising front end for selective surveillance

We present SAR-RAG, an agentic image-retrieval-augmented generation (ImageRAG) framework for automatic target recognition (ATR) in synthetic aperture radar (SAR) imagery. SAR is widely used in defense and security, but visually similar vehicle signatures, speckle, and sensor-dependent effects can complicate interpretation. SAR-RAG combines a multimodal large language model (MLLM) with a vector database of semantic SAR embeddings to enable contextual search over a library of previously observed, labeled exemplars with known target types and associated attributes. During inference, the agent retrieves the most relevant reference images and uses them as an attached “memory bank” to support comparison and reasoning across closely related vehicle categories, improving both categorical identification and quantitative estimation. We evaluate the approach using retrieval quality measures, ATR classification accuracy, and regression of vehicle dimensions, and show consistent improvements over an MLLM-only baseline when retrieval-augmented context is incorporated.

Infrared (IR) detection and tracking technologies are advancing rapidly with the integration of artificial intelligence and machine learning (AI/ML). While these tools enhance capability and automation, they also present challenges in ensuring consistent and transparent system performance across diverse operational conditions. Reliable AI/ML training depends on large, high-quality datasets, yet real IR data are often limited. Synthetic data offers a promising solution, but their influence on model performance and trustworthiness remains uncertain. This work compares AI object detectors trained using the empirical ATR Algorithm Development Image Database (ADID) and a synthetic counterpart developed by Perez, Vanstone, et. al. Leveraging explainable AI (XAI) tools the study evaluates how synthetic data may affect model decisions. Results contribute to a framework for validating AI/ML systems that use synthetic data, promoting greater transparency, reliability, and confidence in future IR sensing applications.

The naval ship infrared signature model and naval threat countermeasure simulator (ShipIR/NTCS) developed by W.R. Davis Engineering Ltd. has undergone extensive validation since its adoption as a NATO-standard and through US Navy Accreditation for Live Fire Test and Evaluation of the DDG-79 (Flight IIA) Aegis Guided Missile Destroyer and Contract Design of the DDG-1000 (USS Zumwalt) Guided Missile Destroyer. Some of the features that make this tool well suited for input to AI / ML applications is its fully-deterministic approach to thermal / EOIR modelling. With the delivery of a turn-key ShipIR thermal and EOIR model of each class of ship, only a few climatic and ship operational inputs are required to construct a full operating EOIR scenario with both transient background and ship signature conditions.This paper will describe the overall framework but also its scene rendering and scenario model updating capabilities. It is hoped that new use case evaluations might result from the presentation of these details to a wider Artificial Intelligence and Machine Learning Community.

Modern automated target recognition (ATR) systems can use a diverse set of modalities to identify targets. Of particular importance are infrared emissions for night-time detection. In this paper, we describe a physically implemented infrared camouflage system that uses a limited number of temperature-adjustable panels that can be placed on a vehicle and adversarially trained to evade ATR by changing the infrared emitted by the panels. Testing with a YOLO ATR with and without the camouflage activated, we consistently degrade detection from 90% to below 52% confidence. This was achievable even when only activating 14 of the 24 available panels. This improves upon previous work by creating a cost-effective solution that is optimized to avoid ATR instead of attempting to modify the infrared pattern of the entire target.

Thermal (infrared) perception is important for robust object detection in low-illumination and poor-visibility conditions, especially in automotive scenes. This work studies parameter-efficient adaptation of the open-vocabulary detector YOLO-World from visible to thermal imagery using two complementary strategies: input-space modality prompting (ModPrompt) and weight-space low-rank adaptation (LoRA). Experiments on the FLIR-IR aligned benchmark with three categories—person, bicycle, and car—show that both methods improve over zero-shot transfer, while their combination achieves the best accuracy (AP50 75.5, AP75 36.9, AP 41.1). LoRA offers the best efficiency among the tested settings, whereas LoRA combined with ModPrompt provides the strongest detection performance at higher computational cost. These results highlight a practical trade-off between efficiency and accuracy for thermal open-vocabulary detection.

Detecting small UAVs in thermal infrared imagery is challenging because targets often occupy only a few pixels, have weak contrast, and appear in cluttered backgrounds. We propose a real-time hybrid detector that runs YOLOv11-s and a multiscale Relative Local Contrast Measure (RLCM) branch in parallel, then fuses their candidate detections at the bounding-box level before lightweight temporal confirmation with a probabilistic data association filter (PDAF). The PDAF associates detections across nearby frames, suppresses isolated clutter responses, and confirms persistent targets. On Anti-UAV410, the proposed method improves AP over YOLOv11-s, with the largest gains on tiny targets, increasing APt from 0.47 to 0.70. On Jetson AGX Orin Industrial, the full pipeline runs in 4.53 ms per frame versus 4.42 ms for YOLOv11-s alone, adding only 0.11 ms latency. These results show that candidate-level fusion of semantic confidence and contrast-based saliency, together with lightweight temporal confirmation, improves tiny infrared UAV detection while preserving real-time embedded performance.