A traffic object classification pipeline based on CNNs

In China, a variety of non-motor vehicles (NM-Vehicle) exist on the road in recent years, some are similar to pedestrian, others are similar to motor vehicle (M-vehicle) in monitor video, which brings a great number of trouble to policeman when case tracking. In this paper, we proposed a traffic object classification pipeline based on state-of-the-art deep convolutional neural networks (CNNs).Unfortunately, there is no traffic object image database which typically includes motor vehicle (M-vehicle), non-motor vehicle (NM-vehicle) and pedestrian images as we know, and CNNs will be suffered over-fitting without enough training images. Therefore, we first construct a large-scale traffic object image dataset with more than 50000 training images. Then histogram equalization was used to compensate the poor illumination caused by the diversity in image scene. Considering the computing time, we finally choose the 3-convlayer network to our cooperative engagement information system (CEIS), and the value of mAP is 93.3%. With the model fusion, the result is improved to 96%.