Learning Based Road Estimation

Abstract: The interest in autonomous driving has vastly increased, leading to a surge in research and development efforts over the past decades. This technology could enhance road safety, alleviate traffic congestion, and yield numerous environmental and economic benefits. A fundamental prerequisite to developing and integrating autonomous driving is to obtain information about the surrounding environment, particularly in terms of detecting lane geometry. Precise lane detection can be achieved in a number of ways, including with the use of deep learning. This thesis investigates deep learning-based lane detection by developing models that predict the center of the ego lane. The problem is treated as a regression problem and is approached by developing, testing, and comparing several model architectures by using mean squared error as metric. Lane marker detection coordinates were used to generate bird's eye views (BEV) used as input to the different models. Four main models were developed. Model 1 was a convolutional neural network (CNN) and Model 2 was the CNN with a one-dimensional input fused to the feature map of the CNN. Model 3 incorporated a long short-term memory (LSTM) after Model 2 to make use of the temporal information in the data. Furthermore, Model 4 was created which had the same architecture as Model 3 but with another loss function. The experimental results demonstrated a consistent decrease in loss with the introduction of each subsequent model. The conclusions could be drawn that the selection of a metric used to train and evaluate networks was essential in the process of developing a relevant model. From the experience gathered during the thesis, a learning-based approach seemed to hold a lot of potential for estimating road geometry.