| Potential safety hazards caused by different types of UAVs are different.Different countermeasures should be taken for different types of UAVs intrusion.Mass data,superior hardware and improved algorithms make convolutional neural networks have better classification performance in computer vision and other fields.The accuracy of the existing UAVs classification methods detected by radar needs to be further improved.The problem is how to effectively use convolutional neural networks to perform deep learning of UAVs’ micro-Doppler features and achieve UAVs classification detected by radar.Therefore,in order to solve the problem,the method is conducted in-depth research which uses convolutional neural network to classify UAVs detected by radar.The research provides technical support for relevant departments that take different countermeasures for different types of UAVs.Firstly,training sample sets are constructed which are based on rotor UAVs’ micro-Doppler features.Rotor UAVs’ simulated radar echo data are carried out micro-Doppler features analysis and extraction.The Merged Doppler Image datasets of rotor UAVs are constructed.Secondly,impacts of training sample sets on UAVs’ classification results are analyzed.Simulated radar echoes data of three types of rotor UAVs are generated.The influences of rotor UAVs’ parameters on micro-Doppler features are analyzed,including blade rotation speed,blade length,blade initial phase,UAVs’ azimuth,UAVs’ pitch angle and UAVs’ radial velocity.Merged Doppler Images(MDI)training sample sets are constructed in many different situations.Goog Le Net(Inception v1)is used to obtain UAVs’ classification results in different situations.Impacts of training sets on the classification accuracy are analyzed,including sample quantity,variation of UAVs’ single parameter,completeness of sample parameters coverage and sampling interval of UAVs’ parameters.According to the analysis results,suggestions that construct training sample sets are given,which are used to guide the collection of measured data in the process of constructing measured training sample sets.Finally,an algorithm of classifying UAVs detected by radar is given which is based on improved Dense2 Net.Simulated radar echoes data of rotor UAVs perform wavelet denoising.The micro-Doppler features analysis and extraction are performed to obtain UAVs’ Merged Doppler Image datasets.The Merged Doppler Image datasets are sent to improved Dense2 Net for feature extraction and classification results of UAVs are obtained.The experiment results show that the improved Dense2 Net performance is better and the classification accuracy is higher.Improved Dense2 Net represents multi-scale features at a finer-grained level and increases the receptive field of each network layer. |
