SegLink: An Advanced Lane Detection Network with Minimal Computational Overhead Using Semantic Segmentation

With the growth of global population and increasing number of road vehicles the driving safety has become crucial. To cater this problem, an efficient lane detection system is proposed by using semantic segmentation-based encoder -decoder convolutional neural networks with reduced filter size. For this, three networks named as SegLink-v1, SegLink-v2 and SegLink-v3 are proposed where features generated at different hierarchical levels of encoder are connected to corresponding decoder layers. Each decoder layer performs convolution on the resultant of corresponding encoder layer and previous decoder layer. In comparison to benchmark SegNet our networks demonstrated improved test accuracies of 1%, 1.01% and 1.5% respectively. Whereas SegLink-v2 and SegLink-v3 networks achieved 43.44% and 75% reduction in computational parameters. Due to their high accuracy and less computation parameters, they may be used for designing efficient driving assistance system.