Adaptive vehicle classification based on information entropy and multi-branch BP neural networks

Vehicle classification is an important and hard process in traffic information collection. Feature extraction and pattern recognition are two of the most important sub-processes in vehicle classification based on the signal curves of the loop transducer. An adaptive classifier, which uses knowledge base and information entropy for feature extraction, and multi-branch BP neural networks (MBBPNNs) for pattern recognition, is proposed in this paper. NMBPNNs are designed according to the divide-and-conquer and modular principles. A N-class problem can be divided into N parallel two-class problems by NMBPNNs, and the limitations of general BP neural net-works in N-class problems will be reduced. Finally. This paper studies the application of the classifier in vehicle classification, and does simulation with some actual data collected from Chinese national highway G107. The simulation results confirm that the classifier will improve the performance and remarkably reduce the training time against other classifiers that use general feature extraction and BP neural networks for pattern recognition.