Research On The Application Of Feature Extraction And Feature Optimization In Vehicle Acoustic Classification

As an essential part of modern wireless sensor network (WSN), target acoustic classification system (TACS) can accurately sense the presence of maneuvering objects, such as vehicles, helicopters and tanks, and estimate their type and quantity. With the collaboration of other modules of WSN, including target positioning, tracking and attacking module, TACS has profoundly changed the way of traditional war, bringing forth a whole new warfare ideology of detection and attack with super intelligence.The thesis has designed and implemented a vehicle acoustic classification system which can distinguish two kinds of vehicles, namely Assault Amphibian Vehicle (AAV) and Dragon Wagon (DW). The system consists of four modules, signal acquisition, signal detection feature extraction and pattern classification. The thesis proposes a new acoustic detecting method based on OS-CFAR(Order Statistic-Constant False Alarm Rate), which can accurately extract real vehicle acoustic signals. In the thesis, the theory of pattern recognition was first introduced, especially the methods of feature extraction and feature optimization. Then several acoustic features were extracted, including short time energy, zero crossing rates, harmonic set, wavelet band energy, and so on. By using class separability measures on these features, a candidate feature bank was formed without some bad and eliminated features. In the final, the genetic algorithm was applied to optimize the feature bank and select the corresponding elements from it, resulting in a subset feature vector.The experiment results showed a great classification performance improvement in subset feature vector, raised from the original 56.7% to 11.7%.