Vehicle-pedestrian Classification Based On Deep Neural Networks

After the discovery of micro-Doppler e ff ect caused by micro-motion of targets, a new source of features has been introduced to radar target classification and recognition, making it possible by means of signal processing to extract the spectrogram of radar echo which can reflect the movement of small parts of targets, while it is very di ffi cult to obtain this information without micro-Doppler e ff ect. There is a great need of automatic classification of pedestrians and vehicles, because the ability of a ground surveillance radar to identify a target can dramatically improve the automation of military equipments, exclude errors of human operators, and give support to a right military decision-making. In this paper, micro-Doppler target detection and classification techniques are systematically studied based on the real world application requirements, the research results obtained are:First, Radar signal processing and micro-Doppler feature extraction methods based on Linear Frequency Modulated Continuous Wave(LFMCW) system are studied. Multi-cycle signal models of a LFMCW system and its beat signal frequency response characteristics are discussed, providing theoretical support for the subsequent signal processing. For this vehicles and pedestrians classification problem, signal preprocessing methods should be able to maximize the retention of micro Doppler signal components and suppress clutter as best as possible at the same time. Therefore CLEAN algorithm was introduced to increase the Signal-to-noise ratio(SNR) of Radar echo. Also we focus on the mathematical model of micro-Doppler e ff ect, by analyzing and comparing the di ff erences between micro-Doppler spectrum of pedestrian and slow vehicle, indicating the feasibility of target classification based on micro-Doppler features.Second, Artificial neural networks(ANNs) algorithms are studied thoroughly. We derived the back-propagation algorithm in detail and discussed the ability of a large-scale neural network model to match any function, and thus brought about over-fitting problems and its corresponding solutions — regularization. We also studied sparse auto-encoder as well as its training methods, and discussed the concepts and superiority of representation learning.Concepts and mathematical models of convolutional neural network(CNN) were also studied. We propose that sparse auto-encoder working with CNN and a classifier can complete the feature learning and target classification in the absence of human intervention based on micro-Doppler spectrum.Third, di ff erent unconstrained numerical optimization methods, which are needed to train a neural network, including the steepest descent method, Newton methods, quasi-Newton methods, limited memory quasi-Newton methods and their simulation tests, are analyzed deeply. And their advantages and disadvantages are discussed, to provide a theoretical basis for a fast training of a large-scale neural network model.Finally, we propose a new ground moving target classification algorithm based on feature self-extractation, whose algorithm procedure is distinct with traditional radar target recognition methods, which rely on hand-designed feature extraction heavily. We also figured out how to work CNN into processing micro-Doppler spectrum, including data normalization and whitening, and ultimately do ground moving target recognition. And experiments are conducted to show that it can e ff ectively separate pedestrian and vehicle targets.