Trajectory Prediction of UAVs: A Deep Learning Approach

Trajectory Prediction of UAVs

Villadsen, Rasmus Erik

4. semester

Mathematical Engineering, Master

2026

2026-05-26

This thesis evaluates how well modern deep learning models can predict the three-dimensional flight paths (trajectories) of unmanned aerial vehicles (UAVs). Unlike much prior work that relies on simulations, the study combines eight real-world datasets. To provide baselines, the study implements simple motion models - constant velocity and constant acceleration - as well as an unscented Kalman filter (a statistical tracking method). As deep learning baselines, long short-term memory (LSTM), gated recurrent unit (GRU), and Transformer architectures are used, and the evaluation is expanded by introducing two architectures, Mamba-1 and Mamba-2. All models are trained with two types of input data: position coordinates and velocity derived from position (first derivative). Training and testing cover different context lengths (how many seconds of past data the model sees: 5 and 10 seconds) and prediction horizons (how far ahead it predicts: 1, 3, and 5 seconds). Model settings are tuned with a grid search. Performance is measured with root mean square error (RMSE), which summarizes average prediction error, and final displacement error (FDE), which measures the error at the final predicted point. Across most combinations of context length and prediction horizon, position-based Mamba-1 models outperform both the other deep learning models and the baselines. The results also show a clear gap between input types: models trained on velocity data are less accurate and have more variable performance.

Dette speciale undersøger, hvor godt moderne deep learning-modeller kan forudsige tredimensionelle flyvebaner (trajektorier) for ubemandede fly (UAV'er). I modsætning til megen tidligere forskning, der bygger på simuleringer, kombinerer studiet otte virkelige datasæt. Som grundlinjer implementeres simple bevægelsesmodeller - konstant hastighed og konstant acceleration - samt et unscented Kalman-filter (en statistisk sporingsteknik). Som deep learning-baselines anvendes long short-term memory (LSTM), gated recurrent unit (GRU) og Transformer-arkitekturer, og evalueringen udvides med to arkitekturer introduceret her, Mamba-1 og Mamba-2. Alle modeller trænes med to typer inputdata: positionskoordinater og hastighed afledt af position (første afledte). Træning og test dækker forskellige historiklængder (hvor mange sekunder af fortiden modellen ser: 5 og 10 sekunder) og forudsigelseshorisonter (hvor langt frem der forudsiges: 1, 3 og 5 sekunder). Modelindstillinger justeres med en systematisk afprøvning (grid search). Ydelsen måles med root mean square error (RMSE), som opsummerer den gennemsnitlige forudsigelsesfejl, og final displacement error (FDE), som måler fejlen i det sidste forudsagte punkt. På tværs af de fleste kombinationer af historiklængde og forudsigelseshorisont overgår positionsbaserede Mamba-1-modeller både de andre deep learning-modeller og baselines. Resultaterne viser også en tydelig forskel mellem inputtyper: Modeller trænet på hastighedsdata er mindre præcise og har mere svingende resultater.

[This abstract has been rewritten with the help of AI based on the project's original abstract]

Trajectory prediction ; Trajectory prediction of UAVs ; Time series forecasting ; UAV ; Kalman filter ; Unscented Kalman filter ; State Space Models ; Gated Recurrent Unit ; GRU ; Long short term memomy ; LSTM ; Transformer ; Attention ; Mamba ; Mamba-1 ; Mamba-2 ; State Space Duality ; Dataset ; Big dataset ; Big data

Download PDF
View record in AAU Student Projects