Abstract:
In Advanced Driver Assistance Systems (ADAS) which has been actively researched in recent years, it’s very important to understand the situation around the vehicle. Espec...View moreMetadata
Abstract:
In Advanced Driver Assistance Systems (ADAS) which has been actively researched in recent years, it’s very important to understand the situation around the vehicle. Especially, since blind spots are dangerous for the driver, it’s necessary to reduce them in some way. In Japan, many road safety mirrors, which are convex mirrors for reflecting vehicles and pedestrians in the blind spot, are installed at intersections, T-junctions and sharp curves with poor visibility. This paper proposes the method to recognize the situation of blind spots using a road safety mirror and an in-vehicle camera. There are two major problems in recognizing images in road safety mirrors. One is that it’s difficult to detect and track objects in mirror images since a mirror image taken by an in-vehicle camera has low resolution. The other is that there is no public database focusing on road safety mirrors and it’s not easy to collect their images. To deal with low resolution, our method extracts moving features from mirror image sequences using DNN instead of detecting objects in each image. To compensate for not enough real-data, we also create synthetic image datasets in which CG-data style is converted to real-data style by image-to-image translation network. Our model is trained with CG datasets of real-data style. Evaluation results using our own dataset show that our method can be adapted to low-resolution images where object detection is not possible.
Date of Conference: 21-25 August 2022
Date Added to IEEE Xplore: 29 November 2022
ISBN Information: