The vehicle classification correct rate of loop induction detector widely used on road is not high. The main reason is the classifier of fixed classification rules cannot cope with the changes of vehicle's complicated models and new vehicle's type classing. Based on the electromagnetic induction characteristic waveform of the vehicle passing through the loop，a new real-time discriminant method for vehicle classification was proposed. The principal component analysis method was used to extract the features. The adaptive resonance neural network algorithm was applied to cluster classification modes，these were dynamically divided into vehicle types then. For new vehicles of unknown vehicle type，new classification modes were added online by semi-supervised learning to adapt to the recognition of new vehicle type. The average correct rate of road real-time vehicle identification experiments of 7 models was 91.3%， and it was increased to 92.5% after adding new mode automatic recognition. In the comparative experiment with Alexnet multi-layer convolutional neural network algorithm， the correct rate of training set and test set were 99.5% and 87.1% respectively， which signified the existence of big differences. The experimental results verified the feasibility of the proposed method to solve the road vehicle identification problem of the change of vehicle mode.