Performance evaluation of semantic segmentation using efficient neural network (ENet) on various traffic scene conditions / Miza Fatini Shamsul Azmi, Fadhlan Hafizhelmi Kamaru Zaman and Husna Zainol Abidin

Shamsul Azmi, Miza Fatini and Kamaru Zaman, Fadhlan Hafizhelmi and Zainol Abidin, Husna (2021) Performance evaluation of semantic segmentation using efficient neural network (ENet) on various traffic scene conditions / Miza Fatini Shamsul Azmi, Fadhlan Hafizhelmi Kamaru Zaman and Husna Zainol Abidin. Journal of Electrical and Electronic Systems Research (JEESR), 19 (1): 18. pp. 135-142. ISSN 1985-5389
Identification Number (DOI): 10.24191/jeesr.v19i1.018
Abstract

Object recognition, object detection, and semantic segmentation are fundamental components of the intelligent vehicle. Recently, there have been various methods proposed to create a reliable and accurate model to provide intelligent assistance to drivers. However, a reliable and accurate model in adverse conditions such as snow, rain, and fog remain a problem for advance driving assistance systems. The methods proposed only effectively solve the problem in a specific condition. Therefore, in this work, we focus on performing semantic segmentation in normal, rainy, foggy, and low light conditions using Efficient Neural Network (ENet) and ResNet18 and subsequently evaluating the trained model’s performance in these conditions. In the experiment, we used a daytime data set from CamVid and synthetically transformed the daytime data set into rainy, foggy, and low light conditions. To verify the accuracy of the proposed method, the Intersection over Union (IoU) is used, and the result is elaborated in the section result and discussion. This approach only performs accurately during daylight. From the experiments, both methods do suffer from various conditions, but the ENet method performs better in certain conditions compared to ResNet18.

Item Details
Edit Item
Edit Item
Downloads & Files
[thumbnail of 52080.pdf]
Text
52080.pdf
Download (802kB)
Indexing & Metrics