SDV-Net: A Two-Stage Convolutional Neural Network for Smoky Diesel Vehicle Detection

Abstract

Smoky diesel vehicle which constantly emit a lot of black smoke is one of the major sources of urban air pollution. It is a timely-responsive, convenient and cost-effective way to detect smoky diesel vehicle utilizing electronic surveillance cameras at both urban and suburban traffic crossings. However, the water stain, shadow and inter-vehicle occlusion significantly hinder the usage of traditional image processing methods in smoky diesel vehicle detection. In this paper, a two-stage convolutional neural network is proposed, codenamed SDV-Net, for smoky diesel vehicle detection. Specifically, an object detection network is firstly presented to predict bounding boxes enclosing the whole outline and back contour of the diesel vehicle, and to locate the regions around exhaust pipes. Secondly, a multi-region convolutional tower network is proposed to conduct a fine-grained classification task on the cropped regions to further determine whether there are smoky diesel vehicle. Several experiments are implemented based on real surveillance images captured by electronic surveillance cameras mounted at different crossroads in Hefei during 2018, and the experimental results explicitly demonstrate the promising effectiveness of our proposed architecture.