A Modification of U-Net Decoder Architecture Improve Performance of Object Detection for Autonomous Vehicles

Autonomous vehicles have made significant advances these days. One of the concerns in its development is the accuracy of object detection on the road to ensure safety in driving. This research solves the problem by modifying U-Net algorithm by developing a segmentation system of objects on the road, specifically on cars and motorcycles. Modifying the decoder by adding a resized input layer to the unit is a novelty in this study. This layer integrates spatial information from the original input into each decoder block through a concatenate operation. To test the effectiveness of the modification, the dataset used consisted of 131 images from streets in the city of Makassar, South Sulawesi. The entire research process, from data collection to model evaluation, aims to measure performance improvement using mean Intersection over Union (m-IoU) and F1-Score metrics. The results showed that the U-Net decoder modification increased the accuracy of detection with m-IoU by 87% and F1-Score by 85%. The findings from this research underscore the significance of algorithmic modifications in enhancing the accuracy of object segmentation for autonomous vehicles.