AI
Multi-target tracking in GPS-denied environments: reliability as an operational defense requirement
A tracking system can show 95% accuracy in a demonstration and lose half of its targets as soon as the environment degrades. On a test bench, the scene is clean, the GPS signal is present, and objects are spaced out. In the field, it is the opposite: jamming, occultations, high target density, and partially blind sensors. The average performance measured in a laboratory says nothing about the reliability of each track at the moment a decision is made.
This is the starting point for TrustalAI's entry into the Defense vertical. The objective is not to detect more. It is to know, with each prediction, whether tracking a target is worth exploiting, even when conditions silently make the AI less reliable.
The real problem: degraded operational conditions, not insufficient AI
Military perception and tracking models are trained on data. In a contested environment, this data no longer resembles the training data.
Electronic jamming deprives the system of reference points. Dense scenes multiply close, moving, and sometimes indistinguishable targets. GPS-denied environments remove the localization anchor upon which most fusion pipelines rely. The concrete result: ghost detections, missed targets, and identity swaps between two crossing objects.
The problem is not that the model is bad. It is that it continues to produce outputs with the same confidence, whether it is within its domain of validity or outside of it. An AI that fails silently is more dangerous than an AI that fails loudly. In defense, this silence has an operational, and sometimes human, cost.
The limitation of learning-based tracking in defense
Learning-based tracking learns a representation of the world from a dataset. This approach works as long as the terrain remains close to the training domain. It weakens as soon as it moves away from it.
Three limitations systematically recur.
First, model drift. A model calibrated for a specific theater, season, or sensor signature sees its reliability drop in another context without signaling it. Model drift is silent by nature.
Second, the dependence on retraining. Adapting a learning-based model to a new environment requires collecting representative data, annotating it, retraining, and revalidating. This cycle is long and costly. It is often impossible in a context where field data is scarce, sensitive, or classified.
Finally, the opacity of the decision. Many systems remain black boxes. They provide an output, not a measure of the confidence to place in that output. Yet, in a contested environment, the useful question is not "what is the estimated position?" but "can I trust this specific track, right now?".
Recent work describes this problem as "silent failure": AI systems that produce confident errors without any warning signal, even in critical production (Betakit / ApplyBoard).
TrustalAI V-Tracking: non-deterministic, learning-free tracking
TrustalAI V-Tracking addresses this limitation through a shift in logic. Instead of learning yet another representation, the product adds a reliability-per-prediction layer to the existing detection flow.
Tracking is non-deterministic: each track carries an estimation of its uncertainty, not just a position. It is learning-free: it works on the existing detection flow, without retraining the model and without accessing its IP. It runs on the edge with a latency of 20 ms, well below the <100ms threshold required for real-time applications.
What this produces, measured on the official product metrics (TRL9):
x2.1 objects tracked simultaneously
81.9% recall, 97% precision
-64.7% localization errors
-96.7% false detections
-46% identity switches
Identity switches are worth pausing on. In a dense scene, two crossing targets can have their identities swapped by the tracker. Over a tracked trajectory, this swap propagates an error downstream. Cutting these switches in half means maintaining tracking continuity where traditional approaches lose it.
On the perception side, TrustalAI Vision plays the same role at the detection level: a plug-and-play reliability block, compatible as a black-box with existing vision models, which attaches an uncertainty measurement to each detection (-80% errors, -83% false positives, 20 ms). Vision and V-Tracking combine: one secures detection image by image, the other secures tracking over time.
A point of honesty: these figures are confidence metrics validated at the product level, in a real operational environment (TRL9). They do not represent published defense field results. The building block is proven; its application to a given theater remains to be qualified on a case-by-case basis.
Per-prediction reliability vs aggregated monitoring: the distinction that matters
Most AI supervision approaches monitor the behavior of a system after execution. They aggregate metrics over batches of decisions, produce dashboards, and signal a drift once it has already set in. This is post-mortem monitoring.
The issue is timing. By the time an aggregated metric reveals a problem, the decision has already been made and the action has already occurred. In defense, the gap between "the system malfunctioned" and "we found out" is paid for on the mission.
Per-prediction reliability reverses this logic. It does not judge the model on its average. For each track and each detection, it measures if that specific output is reliable enough to be followed, before the decision is made. An average performance of 97% can hide the 3% of cases where the system fails, and those are precisely the cases that matter in a contested environment.
In other words: aggregated monitoring tells you how your model behaved yesterday. Per-prediction reliability tells you if you can trust it right now, on this specific target.
EU AI Act compliance and military systems: a secondary issue
The regulatory question arises as soon as critical AI is discussed. For strict defense applications, it is lighter than elsewhere.
The EU AI Act (Regulation 2024/1689) excludes from its scope AI systems developed or used exclusively for military, defense, or national security purposes (Article 2, paragraph 3). A strictly military system is therefore out of scope: the obligations of the "high-risk" regime do not apply to it as long as it remains confined to these uses. The pressure that weighs heavily on ADAS, biometric video surveillance, or industrial quality control is marginal here. However, these systems remain regulated by other means: national law, export controls, international law.
Dual-use nuances this observation. Recital 24 of the regulation specifies that a system designed for military use falls back into scope as soon as it is used, even temporarily, for civil, humanitarian, or law enforcement purposes. A product placed on the market for both an excluded use and a non-excluded use falls under the regulation for this second use, and its provider must then ensure compliance. The same perception block sold in civil and military versions therefore keeps its civil line within the scope of the AI Act.
Added to this is an expectation of reliability that goes beyond the regulatory framework. Emerging certification standards for embedded AI, such as the UL 3115 standard focused on robustness and maintaining a "human in control" (Fortune), do not derive from the AI Act but establish a requirement for proof. Under the high-risk regime, being able to replay a decision (model version, data used, controls applied, human intervention points) becomes an acquisition criterion as much as an obligation (The Recursive).
TrustalAI contributes to this reliability documentation without certifying compliance: certification remains the responsibility of competent authorities and notified bodies.
What per-prediction reliability changes in the field
There remains the question that interests a program manager: what does this change once deployed?
Tracking where the uncertainty is known at each moment allows for prioritization: process reliable tracks first, put non-reliable ones on hold, and alert when the system goes out of its domain of validity. Fewer identity switches mean tracking continuity is maintained on high-density moving targets. Fewer false detections mean fewer resources committed to targets that do not exist, reducing collateral risk.
The learning-free nature also changes deployment economics. No classified data collection, no retraining cycle for each new context. The reliability layer is added to the existing system, plug-and-play, without touching the underlying model.
Finally, there is a sovereignty issue. According to SaferAI, Europe is more exposed than the United States to sectors where AI cannot be adopted without strong guarantees of reliability (SaferAI, March 2026). For defense, relying on an external provider for the reliability block of its critical systems is not neutral. A European reliability layer meets this requirement. The debate over the independence of AI providers from public procurement bodies is, moreover, already open (Al Jazeera).
Reliability does not start after the decision. In a contested environment, it must be visible at the moment the decision is made.
FAQ
What is a GPS-denied environment and why does it complicate AI tracking?
A GPS-denied environment is an area where the geolocation signal is jammed, degraded, or unavailable. Most fusion pipelines rely on this anchor to correlate sensors. Without it, multi-target tracking accumulates localization errors and identity swaps, especially in dense scenes.
How does learning-free tracking differ from learning-based tracking?
Learning-based tracking learns a representation from a dataset and must be retrained to adapt to a new context. Learning-free tracking, like TrustalAI V-Tracking, works directly on the existing detection flow, without retraining or accessing the model. It therefore adapts without requiring data collection and annotation cycles.
What is per-prediction reliability?
Per-prediction reliability measures, for each detection or track, whether that specific output is reliable enough to be exploited, in real-time and before the decision is made. It differs from aggregated monitoring, which evaluates the model retrospectively on an average and detects problems after they have already occurred.
Does the EU AI Act apply to military systems?
Not for strictly defense applications. The EU AI Act (Regulation 2024/1689, Article 2, paragraph 3) excludes from its scope systems used exclusively for military, defense, or national security purposes. These systems fall under other frameworks: national law, export controls, international law. The regulation can apply in case of dual-use: a system also used, even temporarily, for civil or law enforcement purposes falls back into its scope for those uses.
Sources
Share
Related articles
