Lane-aware image enhancement for lane detection in rain (part A)

Abstract

In recent years, due to the continuous development of science and technology, artificial intelligence technology is also under continuous development. The application of AI has led to greater development in the research of autonomous vehicles. Autonomous Vehicles can locate themselves in the environment and recognize the surrounding environment with good environmental fusion cognitive ability. Vision is commonly adopted in Autonomous Vehicle for its rich structural and contextual information on the environment. Several deep learning algorithms may be applied to the visual feedback to easily detect various traffic objects and hazards. However, under rainy conditions, the effectiveness of these detection algorithms may drop due to noise contributed by rain particles. Rain artefacts distort shape of objects, Rain also makes the ground reflect light. Therefore, the wrong classification and wrong location may be caused by rain streaks.

In order to solve this defect, the previously method that can be used to improve detection in rainy conditions is using vehicle rear lights to localize the vehicles. However, the limitation of this approach is that reflection against the wet floor surface resulting in false positive detection. Facing these difficulties, we consider using attention maps to assist the network to enhance the detection accuracy.

In this project, we mainly developed two methods to detect objects on the road, such as trucks, passenger cars, road signs, etc. The first method is to use Faster R-CNN algorithm to test rainy images; the second method is to use SDD algorithm to test dataset. These two methods have their own advantages and disadvantages. Among them, the training speed of Faster R-CNN is slower but the accuracy rate is higher than that of SSD algorithm. The SSD algorithm still needs improvement for the detection of small targets.

In the future, we also hope that this shortcoming of SSD algorithm can be improved. At the same time, we may be able to use keypoint maps to test images, such as CornerNet, CenterNet and other algorithms.