Classification of targets by using FMCW radar data with machine learning methods

Nowadays, object classification has many applications field, such as defense, surveillance, robotics etc. This situation brings so many solutions for problems in traditional classification methods. For example, many values such as speed, Radar Cross Section (RCA), and phase change are taken into account when classifying objects from radar data. However, with usage of this method has bad accuracy in case of similar RCS and speed targets. For example, in case of birds and drones their size and speeds are similar in many cases. Because of this, classification of bird and drone has poor accuracy when using traditional classification methods (Thresholding etc.). In case of drone-bird classification, those algorithms have lack of accuracy. Since, size of a drone and birds are nearly same, RCS classifications cannot work properly. Considering the symmetrical and asymmetrical conflicts and wars taking place today, it is possible to observe that the use of drones is increasing and can be used effectively in these areas. The fact that drones can be easily obtained and easily manufactured can cause this situation. The importance of classifying and distinguishing between drones and birds is important both on the battlefield and in protecting high-value sites and infrastructure from terrorist attacks. For example, airports, military bases, military convoys, government buildings, nuclear power plants, power stations, dams, etc. For this reason, it is important to distinguish these drones using different sensors and different methods. In order to realize this discrimination with high performance, this study is used a real radar sensor measured In-Phase Quadrature (IQ) data spectrograms to observing performance of Convolutional Neural Network (CNN) algorithms. In this thesis study, using micro-Doppler traces of IQ data obtained from FMCW radar, classical machine learning methods and different deep learning architectures have been used to classify different drone types and to distinguish drones from birds. To conduct this study, two datasets is used, first dataset consist of 6 drone classes, 2 human classes with different movements, 6 different bird classes and 1 corner reflector class were used. In second dataset, only drones and bird measurements were used. As a result of this study, high accuracy rates were obtained with end-to-end CNN architectures and Pre-trained CNN architectures. With the CNN architecture proposed in this study, the highest accuracy value 98.04 % was obtained when features were extracted and classified with Support Vector Machine (SVM) with the usage of nine class dataset. With the usage of second dataset which is include only two class (Birds and drones) the accuracy is 99.4 % with the same CNN feature extraction and SVM method use.