Multi-Layer Perceptron Based Small Uncrewed Aerial System Trajectory Classification

Current vision-based detection, tracking, and classification algorithms for Advanced Air Mobility airspace management face significant challenges when attempting to classify detections or tracks. The sparse feature presence in small detections (as small as 5x5 pixels) poses issues for image-based classifiers, as traditional convolutional neural networks struggle to extract meaningful information from these patches due to the limited resolution. Therefore, an alternate, scale-invariant solution is required. This work proposes the development of a trajectory-based binary classification algorithm to distinguish small uncrewed aerial systems from other detections, such as birds or general aviation aircraft. This classifier utilizes a 12-element, track velocity-based feature space captured over a five second window. Various machine learning techniques are evaluated for this application, including six classical machine learning approaches alongside three multi-layer perceptron architectures. The dataset used in this study was collected at NASA's Langley Research Center in July 2025, using two ground-based sensing nodes to capture three multirotor vehicles across multiple test configurations and acquisition days, resulting in approximately 14,000 manually labeled feature vectors. A comprehensive grid search of model hyperparameters and input formats was carried out, resulting in the training and evaluation of over 10,000 model configurations. The experiments revealed that the deeper (6 and 8 layer) multi-layer perceptrons significantly outperformed classical techniques, with a maximum mAP of 0.8586 and a runtime of approximately 1.1μs per feature vector. This work demonstrates that neural network-based trajectory classification techniques are a viable method for aerial track classification due to their high performance and low computational requirements.