TMNet: A Lightweight Semantic Segmentation Based Lane Detection Network for varying Light Conditions

Driving safety is becoming increasingly critical as the global number of vehicles continues to rise rapidly. There is a need to improve the driving assistance systems to ensure road safety. In this research we have proposed a robust and light weight semantic segmentation-based lane detection network. This network can detect lane area on roads under varying lighting conditions during day and night times. For this, machine learning based algorithm is applied for tone mapping replacing conventional signal processing methods. Two networks are proposed with reduced model size and computation parameters. Results indicate that tone mapped network TMNet demonstrated improved performance than its naive lightweight network. For TMNet true positive rate increased by 1.6% with 88.65% reduction in training parameters and 98.8% reduction in model size in comparison to its benchmark SegNet algorithm. Furthermore, training and test accuracies are comparable to SegNet. Owing to its lightweight, the proposed network can be easily deployed in embedded devices for real timeline detection in driving assistance systems.