Drivable Area Detection Method in Dark Unstructured-Roads Based on CNN Data Fusion With Surface Normal Estimation

Vehicle perception of roads is a critical foundation for intelligent transportation systems. While most existing research on intelligent transportation focuses on structured roads, the identification of drivable areas in complex dark-field environments remains challenging and underexplored. To address this issue, this paper proposes a drivable area detection network based on the fusion of images and point clouds. Specifically, a convolutional neural network (CNN) architecture augmented with a surface normal estimator (SNE) model is proposed. This approach significantly reduces resolution while effectively extracting and fusing features from multiple sensors. The proposed model is trained and validated using the Off-Road Freex Detection (ORFD) dataset, demonstrating its effectiveness and accuracy through comprehensive performance metrics. Finally, real-world vehicle tests are conducted to evaluate the drivable area detection system. These tests involve real-time data collection, image enhancement, and dense upsampling of point clouds. The processed multimodal data are fed into the detection network in real time for synchronous training and learning. The detection and recognition of unstructured drivable areas in a real dark environment perception platform confirm the efficacy of the proposed method.