Classification between birds and UAVs using recurrent neural networks

Abstract

In this thesis, we investigate recurrent neural networks (RNN) for classifying birds and drones from radar tracks. The data set consisted of 425 drone tracks and 4000 bird tracks, and included all features measured by the radar system. On this data set, we have empirically evaluated whether it is better to use Long Short-Term Memory (LSTM) units or Gated Recurrent Units (GRU) and how the sequence length affects performance, for both filter based classification and smoothed based classification where the tracks were re-sampled for a regular sampled interval of the targets maneuver. Lastly, a hybrid architecture consisting of multiple parallel GRU networks which is connected to a naive Bayes classifier was implemented and tested on the data. Our results show that classification is possible with an accuracy of around 90% for filter based classification using a GRU architecture and approximately 88% when classifying by smoothed flight-paths using the hybrid architecture. This thesis has shown that it is possible to distinguish drones from birds with recurrent neural network architectures, using their detection history alone. For further improvement of the classifier, more domain knowledge and a revise of data set labels is needed in the case of the hybrid classifier.