Optimisation of semantic segmentation algorithm for autonomous driving using U-NET architecture

In autonomous driving systems, the semantic segmentation task involves scene partition into numerous expressive portions by classifying and labelling every image pixel for semantics. The algorithm used for semantic segmentation has a vital role in autonomous driving architecture. This paper's main contribution is optimising the semantic segmentation algorithm for autonomous driving by modifying the U-NET architecture. The optimisation techniques involve five different methods, which include; no batch normalisation network, with batch normalisation network, network with reduction in filters, average ensemble network, and weighted average ensemble network. The validation accuracy observed for the five methods were 90.28%, 91.68%, 89.80%, 92.04%, and 92.21% respectively. By reducing the filters in the network, the computation time reduces (Epoch time: 1 s 64 ms/step) as opposed to the typical (Epoch time: 4 s 260 ms/step), but the accuracy reduces. The optimisation techniques were evaluated for metrics like mean intersection over union (IoU), IoU for class, dice-metric, dice_coefficient_loss, validation loss, and accuracy. The dataset of 300 images used for this paper's study was generated using the open-source car learning to act (CARLA) simulator platform.