Exploiting Pre-trained Encoder with Receptive Fields and Squeeze-Excitation module for Road Segmentation

Abstract

Autonomous vehicles will decrease the number of accidents on the road caused by human error. Intelligent vehicles have traditionally advanced in a step-by-step manner. These developments boost the automation scene in vehicles by incorporating systems that facilitate the driver in maintaining a constant speed, adhering to a lane, or transferring control over vehicle and driver. Autonomous vehicles must have a thorough understanding of their surroundings. As a result, object detection and road scene segmentation are critical in navigation for recognizing the drivable and non-drivable areas. Towards the development of the completely automated framework for road scene segmentation, we propose an RFB-SELinkNet that utilizes the SEResNeXt model as a feature extractor and receptive field block (RFB) with squeeze and excitation (SE) module for better feature representations. Our proposed framework outperforms D-LinkNet, Eff-UNet, and other state-of-art models. According to the experiments, the proposed model achieves 0.698 mloU and produces good segmentation outcomes on the validation set of the India Driving Lite (IDD Lite) dataset.