Comparison of Water Detection Models for an Off-road Unmanned Ground Vehicle

Abstract

Water hazards can cause unmanned ground vehicles (UGVs) to become stuck or break down during an autonomous mission, damage electronic components and sensors, and require costly repairs or replacements, making it crucial for UGVs to identify water hazards in real-time, determine secure path around them, or reduce their speed when appropriate to cross them safely. This thesis proposes a water detection system for UGVs in off-road environment. The proposed approach combines convolutional neural networks (CNNs) with transfer learning, leveraging their capabilities for effective water detection. The thesis includes a comprehensive review of traditional sensor-based methods and recent deep learning-based techniques. Real-world data collected in off-road environments are utilized to evaluate the proposed approach, and the method achieves a 0.50 Mean-IoU score and 92.74% accuracy on the test dataset. We also include a comparative analysis of the method with a previous deep learning-based semantic segmentation method for water detection. The comparison provides insights into the relative strengths and weaknesses of these approaches for water detection in off-road environments. Overall, this thesis provides valuable insights into the use of deep learning for semantic segmentation in challenging environments.