Ground Target Recognition In Fiber Sensor Network

In this paper, the problems about ground target detection, classificationin fiber seismic senor network are discussed. The unattended ground sensor(UGS) system was developed in military for ground target detection, classificationand tracking in battle field surveillance. The fiber seismic senors have a lotof advantages such as high sensitivity, electromagnetic resistance and networkorganizing easiness. The fiber sensor network has been an important complementfor ground sensor network. In the battle field, the movement of ground targetsuch as personnel, tracked vehicle or wheeled vehicle motivates the seismic wavesthat propagating along with the ground surface. The seismic signal received bythe seismic senor can be analyzed for target classification.First, the generation and propagation of seismic signal caused by movingtarget is discussed. An detection method of personnel footstep is discussed ac-cording to its impulse nature. Then a model of vehicle vibration when movingis built that can be used to analyze the signal characteristic. The propaga-tion model of seismic wave is more complicated to be analyzed. In propagationprocedure, the amplitude of the signal is attenuated, and the frequency has dis-persion. An improved constant false alarm rate (CFAR) detection method basedon adaptive threshold is used to decrease the false alarm in high backgroundnoise. Second, the feature extraction and selection of seismic signal is discussed.The traditional feature extraction methods in time domain, frequency domainand time-frequency domain are discussed, and an united multi-feature method isproposed. The principle component analysis (PCA) is used to reduce the dimen-sion of the feature vector. According to real-life ground target signal acquired inthe field experiment, a large feature data set of di?erent targets is built. Third,the ground target classification method based on statistical learning theory andsupport vector machine is discussed. The traditional classifiers such as Bayes clas-sifier, neural network have some disadvantages when the sample number is small.The comparison experiment shows that the support vector machine (SVM) clas- sifier outperforms the other method, and have bright future in this application.At last, the problems about multi-sensor and multi-target detection and recog-nition are discussed. The decision fusion strategy based on DS evidence theoryis used to improve the overall performance of sensor network. The multi-targetmixed signal processing is a di?cult problem to be tacked in sensor network. Weuse the independent component analysis (ICA) to separate the mixed signal, andthen to classify them separately. The emulation experiment shows that the ICAis an e?ective method.