Automatic classification of air tracks using raw video from a conventional surveillance radar

Abstract

Detecting the type of an airborne vehicle, such as an air plane or helicopter, spotted by a radar, can be very time consuming and expensive if done by examining sensor outputs. This might cause a problem if a potential threat has been detected since the type, and thereafter identity, of a vehicle must be determined rapidly. This thesis investigates the possibility of classifying jets and helicopters from raw video, collected by a conventional surveillance radar, using a neural network. The raw video is transformed from the time domain into the frequency domain using Fast Fourier Transformation and then plotted as either spectrograms or range doppler plots. In order to remove noise, the plots are filtered using the unsupervised algorithm k-means clustering. Different neural networks such as convolutional neural network (CNN), recurrent neural network (RNN) and convolutional recurrent neural network (CRNN) are then trained and evaluated on the plots remaining after filtering. The plots are treated as images by the network. The results show that the CNN model trained on range doppler plots performs the best with a test accuracy of 90% and a validation accuracy of 83%. Our conclusion from this work is that it seems to be possible for a CNN to distinguish jets from helicopters when it is fed with images of Fast Fourier Transformed raw video, but some more research has to be done. Mainly, the model must be trained and tested on more data. The contribution this thesis provides is not only that it seems to be possible to classify jets and helicopters from raw video data in almost real time, depending on the performance of the computer being used, but also that it seems to be possible to receive a great accuracy when using a CNN for solving the task.