Sun-Glare region recognition using Visual explanations for Traffic light detection
Keisuke Yoneda, Naoki Ichihara, Hotsuyuki Kawanishi, Tadashi Okuno, Lu Cao, Naoki Suganuma
Abstract
In order to achieve automated driving on public roads, image-based recognitions such as object detection and traffic light detection are significant technologies to understand surrounding road situations. However, the automated vehicle might be faced severe situations for sensing due to heavy sunshine. If the image captured by the onboard camera is overexposed, the image information will be lost and there is a risk of false detection. In particular, in traffic light detection where the acquisition of color information is essential, if the front traffic light and the sun overlap during intersection driving, it will not be possible to make an appropriate intersection approach judgment. Therefore, this paper developed the method to recognize sun-glare regions in the image using visual explanations of Convolutional Neural Network (CNN). The CNN outputs an attention map using the Grad-CAM method, and then the global direction of the sun-glare region can be estimated by time-series processing. The developed method contributes to implementing robust image recognitions by estimating the direction in which visibility is reduced by sunlight such as direct sunlight and reflected light from buildings.