Ground Target Recognition Algorithm For Smart Sensor Surveillance Networks

Unattended ground sensor technology used in Smart Sensor Network for remote battlefield surveillance applications requires state-of-the-art algorithms to address the unprecedented challenges faced in detecting and classifying ground targets. The primary goal of this study is to successfully recognize the type of ground vehicles using their acoustic signatures recorded by unattended passive acoustic sensors. The objective is to broadly classify the vehicle into tracked and wheeled categories, and to further identify the vehicle type within these categories. Algorithms for feature extraction, classifier design and multiple targets classification are the main content in our research. The main content and results of study involved in this paper are as fellows.(1) Feature extraction and selection algorithms for ground target recongnition are discussed deeply and systemic. Firstly, physical model for the gerneration and transmission of acoustic singals from a moving ground target is presented, and several feature extraction algorithms using for ground target recognition are presented and summarized. Acording to the characteristic and difficulty of target signals, a feature extraction algorithm based on Wavelet Packet analysis is dicussed further, and the measure and serching alogorithms based on discriminating power of features are reseached to improve the performance of recognition. As a result, an optimized feature extraction algorithm are presented and evaluated by real-world signal datasets.(2) Considering the features that are extracted from the acoustic measurements are time-varying and contain a lot of uncertainties due to variations of the environmental conditions (e.g., terrain and wind), a novel recognition system based on Computational Intelligence is proposed. After train samples have been filtered based on Rough Sets theory, a classifier based on Rough Neural Network is used to recognition type of ground target with varied input in time domain since it has ability to solve the qualitative problem. Experiment results show that high performance improvement is achieved using new algorithm.(3) Considering the training samples are unbalanced among different target categories, and the false accept rate of recognition system using classical classifier is high, general algorithms of one-class classifier are discussed. Firstly, according to the Support Vector Data Description (SVDD) one-class classifier is puzzled with the computation of large scale data, difficult to converge and short of robustness, a novel two-steps training algorithm is proposed. The SVDD model is trained from the parameters of GMM in the first step to extract the critical training samples, which are used to train the one-class classifier in the second step. Experiment results show that the proposed training algorithm is much effective than the traditional one. Secondly, a novel one-class classifier based on Biomimetic Pattern Recognition theory and Principal Curves Analysis is presented, experiment results show that...