Adaptive Vehicle Classification Based on Information Gain and Multi-branch BP Neural Networks

Vehicle classification is difficult yet important in traffic information collection based on the signal curves of loop transducer, and feature extraction and pattern recognition are both key processes in automatic vehicle classification. To reduce the information redundancy of feature set, and adapt the classifier to different application circumstance, an algorithm, which employs information gain for feature selection, is proposed. To avoid the limitations of general BP neural networks in multi-pattern recognition, such as divergence and so on, a multi-branch BP neural network, which divides an N-class problem into N parallel two-class problems, is also presented. Furthermore, experiment using actual data collected from Chinese national highway G107 is carried out, and its results show that the classifier based on information gain and multi-branch BP neural networks improves the performance and reduces the training time significantly.