Two-Stream Neural Network Fusion Model for Highway Fog Detection

The real-time detection of weather conditions on highways has a significant impact on high-speed traffic safety. However, the weather forecast reports weather conditions over a wide range of areas only, which cannot meet the demand of real-time detection of weather conditions in various sections of high-speed traffic. Therefore, we present here a two-stream neural network fusion model for fog detection, which detect current weather condition for the surveillance area automatically. This model is based on a dual branches of deep neural networks, which integrates visual depth maps and dark-channel images for fog detection. These two modalities of features are discriminative in representing the pattern of fog and extracted from the surveillance video frame. The intermediate scores produced by the neural networks are fed into a mean fusion layer for the final prediction. To comprehensively evaluate the performance of our algorithm, we built an Express Way Fog Detection dataset （EWFD）, which covers highway scenes across multiple provinces of China. A variety of highway weather conditions are contained in the EWFD. We conducted a comprehensive analysis and comparison experiment on the EWFD dataset. The results of the experiment also demonstrate that the two-stream neural network fusion model proposed here achieved an accuracy of 93.7%, which is a more than 10% improvement compared to the state-of-the-art classification method ResNet-101.