Weather Condition Classification in Vehicle Environment Based on Front-View Camera Images

Abstract

The current environmental conditions should be monitored during autonomous driving since the different weather conditions can have a different impact on implemented sensor system or on the efficiency of the implemented control system. In this paper, the classification of weather conditions in the vehicle environment is based on images captured by a front-view camera, which are further processed by the simple Convolutional Neural Network (CNN). For model development purposes, training and validation data sets were created from two sources: the BDD100K database and by extracting frames from the collected video sequences. The solution implements an additional mechanism to filter out false predictions based on a circular buffer. The proposed solution achieves the F1 measure of 98.3% for the entire test video frames data set, where it achieves the best results in snowy weather detection (Precision of 100%, F1 of 100.00%) and the worst in foggy weather detection (Precision of 97.25%, F1 of 98.00%).