Efficient lane marking detection using deep learning technique with differential and cross-entropy loss

Abstract

Nowadays, researchers are incorporating many modern and significant features on advanced driver assistance systems (ADAS). Lane marking detection is one of them, which allows the vehicle to maintain the perspective road lane. Conventionally, it is detected through handcrafted and very specialized features and goes through substantial post-processing, which leads to high computation, and less accuracy. Additionally, this conventional method is vulnerable to environmental conditions, making it an unreliable model. Consequently, this research work presents a deep learning-based model that is suitable for diverse environmental conditions, including multiple lanes, different daytime, different traffic conditions, good and medium weather conditions, and so forth. This approach has been derived from plain encode-decode E-Net architecture and has been trained by using the differential and cross-entropy losses for the backpropagation. The model has been trained and tested using 3,600 training and 2,700 testing images from TuSimple, a robust public dataset. Input images from very diverse environmental conditions have ensured better generalization of the model. This framework has reached a max accuracy of 96.61%, with an F1 score of 96.34%, a precision value of 98.91%, and a recall of 93.89%. Besides, this model has shown very small false positive and false negative values of 3.125% and 1.259%, which bits the performance of most of the existing state of art models.