Vision-Based Vehicle Detection in Foggy Days by Convolutional Neural Network

Abstract

Vehicle detection plays an important role in advanced driving assisted system and autonomous driving system. However, the existing vehicle detection methods are not robust in harsh environments, especially in foggy environment. To solve this problem, a vision-based vehicle detection structure using convolutional neural network is presented to detect the vehicle in foggy days. In our vehicle detection structure, a pair of encoders and decoders is used to estimate atmospheric illumination and transmissivity, and to establish the defogging image firstly. And then, the vehicle detection is implemented by a proposed vehicle detection method which predict the left-top key point as well as the right- bottom key point of the vehicle, thus get the bounding box of the vehicle. To verify the effectiveness of the new method, a data set based on the video generated from PreScan simulation platform is set up. And the new vehicle detection method is tested in multiple scenarios such as left turn, right turn, uphill, downhill. Experimental results show that our vehicle detection structure can effectively detect vehicles in foggy days.