Feature Extraction And Recognition Of Radar Signals

The radar is originally a military surveillance telemetry system for the target detection, location and distance measurement. With the development of radar technology, its application is no longer confined to the military field, but has been widely used in all aspects of life and production. However, a problem is still not well solved, namely how to identify targets from the data. For such problem, radar signal feature extraction and recognition is investigated in this paper from both one-dimensional waveform and two-dimensional profile data.For one-dimensional waveform data, clutter filtering, smooth filtering are used in the preprocessing phase to denoise and enhance the stability of follow-up features. Width, area, fractal dimension and coding entropy, along with wavelet decomposition are used to extract features from single echo. Statistical methods are applied to reduce the fluctuations among echoes, and singular value decomposition is used to characterize the relation between echoes. The echo group as a sample and the neural network as a classifier, classification for targets achieves a good result.For two-dimensional profile data, target automatic segmentation, binary, morphological filtering and boundary extraction are fulfilled in the preprocessing phase. Fourier descriptor and chain code are applied for feature extraction and a new method based on Hough transformation is proposed to avoid the drawback of the previous methods. The results of the three types of features are acceptable using the neural network as a classifier. Due to the limitation of the three methods, a new multi-feature synthesized classifying network is designed to improve the stability of the features. From the result it's clear that the performance of the synthesized network is obviously better than the single network, and the characteristics within-class and between-class are described completely.It's concluded that radar signal feature extraction and recognition from waveform and profile data is feasible and the methods proposed are valid.