The research presents a robust method to categorize nighttime traffic by thermal images. The objective of the research is to identify and separate vehicles from unimportant thermal objects of nighttime traffic. Various thermal features of a vehicle comprising windscreen, engine heat and elsewhere, are depicted as different intensities in each thermal image. Moreover, the relation between the three features can be illustrated as a semi-Gaussian statistical distribution of data, which differentiates one type of vehicle from another. To recognize a vehicle type, there are three operating modules proposed in this research, comprising vehicle-thermal features detection, semi-Gaussian evaluation and decision, respectively. Initially, the detecting process finds a vehicle engine heat-feature by subtracting the road's intensity, selecting a suspected area to cover the vehicle's windscreen feature. Secondly, semi-Gaussian thermal distribution model is implemented to evaluate similarity between the suspected area and the model. Eventually, the decision is made based on the similarity to confirm whether it is a vehicular type or not. Experimentally, using four types of vehicles consisting of a car, van, truck and thermal object, the accuracy of identifying is extremely high.