Road Sign Instance Segmentation By Using YOLACT For Semi-Autonomous Vehicle In Malaysia

This paper applies the object segmentation method to road sign recognition by using You Only Look At CoefficienTs (YOLACT). YOLACT achieves speeds higher than 30 fps in real-time with high performance in terms of precision and reliability. However, YOLACT performance is affected by the different situations in the image such as lighting, weather, and different angle on the traffic signs. This research trains and applies the image preprocessing on different angles (such as 90 degrees left and right) and environments to increase the recognition performances. This is important to ensure that the method used can be safely executed in an autonomous vehicle. For preprocessing, the backbone network uses ResNet-101 and is used in the YOLACT system to verify the performance of autonomous vehicles. Four types of traffic signs are used to validate our method. The result shows all traffic signs were successfully identified, with no mislabeled and accuracy exceeding 95%.