| In recent years, object recognition based on data level and feature level becomes a hot topic. Traditional for target recognition based on data level large amount of calculation and recognition effect is poorer. The target recognition based on feature level with the method of remove redundant data, can both retain more effective data and information, and greatly reduces the computational complexity and has higher accuracy. This paper is based on the background of the security enclosure of South to North Water Diversion project.Generally, the object recognition methods in the security enclosure can be classified into 3 levels:data level, fusion level and decision level, based on when the fusion is applied. Corresponding, the corresponding multisensor system is composed by 3 models:data-fusion model, structure model, math model. In this paper, we mainly studied the last two models, and proposed two modified algorithm for pattern recognition:(1) A new classification algorithm combining multilevel threshold classifier on the basis of kurtosis algorithm:Firstly, we extracted eigenvalues from original data, such as mean values, variances and kurtosis. Then, multilevel threshold classifier was introduced to decide which pattern the signals belonged to. It was proved that the classification algorithm leads to ideal result in engineering in short distance (about 100 meters) with rapid computing speed.(2) A new classification algorithm combining wavelet packet decomposition and BP neural network:This algorithm was to overcome the arbitrary decision when setting the threshold. Firstly, the feature level was extracted from the original data by wavelet packet decomposition, then BP neural network was employed to perform the feature classification based on the feature level..In the 100 m distance to pat, climb, swing, trust in four anomaly pattern recognition rate reached 95%.Likely, this algorithm achieved better performance than traditional method and served better robustness. |
PDF Full Text Request