Coupled Multivehicle Detection and Classification With Prior Objectness Measure

Abstract

Vehicle recognition plays an important role in traffic surveillance systems, advanced driver-assistance systems, and autonomous vehicles. This paper presents a novel approach for multivehicle recognition that considers vehicle space location and classification as a coupled optimization problem. It can speed up the detection process with more accurate vehicle region proposals and can recognize multiple vehicles using a single model. The proposed detector is implemented in three stages: 1) obtaining candidate vehicle locations with prior objectness measure; 2) classifying vehicle region proposals to distinguish the three common types of vehicles (i.e., car, taxi, and bus) by a single convolutional neural network (CNN); and 3) coupling classification results with the detection process, which leads to fewer false positives. In experiments on high-resolution traffic images, our method achieves unique characteristics: 1) It matches the state-of-the-art detection accuracy; 2) it is more efficient in generating a smaller set of high-quality vehicle windows; 3) its searching time is decreased by about 30 times compared with the other two popular detection schemes; and 4) it recognizes different vehicles in each image using a single CNN model with eight layers.