Lane Detection for Autonomous Driving: Conventional and CNN approaches

In this talk, Tesfamichael Getahun, a member of ACCESS Laboratory, will present lane detection techniques as one of the fundamental components in advanced driver assistance systems (ADAS) in semi-autonomous vehicles in the form of lane keep assist or lane departure warning. LAne detection is laso one of the key enablers for autonomous driving for it can be used as part of the lateral and longitudinal control system or for localization. Lane detectors determine lane boundaries by extracting the edges and color of paintings on the road. However, the paintings on the road may not always be visible for various reasons such as shadows, aging, road texture changes and other environmental factors which makes vision-based lane detection a challenging problem. In this seminar, Tesfamichael presents his recent results on reliable lane detection to handle some of those challenges. In particular, he presents one of his developed lane detectors which relies on conventional image processing techniques to enhance and extract features from the image stream which utilizes road geometry parameters like lane width and lane marking thickness. He also discusses a lightweight deep neural network approach for feature extraction without the need to manually tune thresholds and other parameters as the network learns them indirectly from training data.