Mathematical programming based solutions for autonomous vehicle in adverse weather

Autonomous vehicles are robots that can navigate by sensing the traffic flow, pedestrians and other vehicles around them without the need for any driver or intervention, thanks to their software and hardware. They do this detection process over the data obtained from sensors such as radar, lidar, stereo cameras. Afterward, they process the obtained sensor data and participate in the traffic flow. However, one of the biggest obstacles to autonomous vehicles being a part of our daily lives is the dramatic decrease in sensor performance in bad weather conditions. In this study, the object classification problem on images obtained from a stereo camera sensor in an autonomous vehicle. In order to solve this problem and obtain comparative results, two separate datasets containing images of the same real-life scenario taken in both sunny and rainy weather were used. A deep neural network-based autoencoder model is proposed to clean the images in the data set, which were taken in rainy weather conditions and deteriorated due to raindrops adhering to the lens of the stereo camera. The obtained results show that the proposed autoencoder improves the object classification performance.