2020Research papers

Summary

Reducing accident risks due to driver's visual workloads is widely expected in recent traffic environment. In this paper, we describe a novel method to classify complex traffic scenes tending to increase the workloads from scattering visual saliency. Classification accuracy is improved by key methods based on a visual saliency predictor and a scene classifier. The predictor is improved by training an encoder-decoder CNN model with reducing pooling layers and with three loss functions. The classifier is improved by training a CNN model using the encoder outputs along with the predicted saliency. In the scene classification tests using our experimental system and SALICON dataset, the proposed method is effective in comparison with baseline methods such as simple VGG classifiers.