YOLO-Based Terrain Object Detection for Autonomous Rover Navigation on Mars

The challenge of navigating planetary terrains autonomously is a significant problem for navigation systems, as they depend on perception tools to determine how well to navigate different terrains in extreme conditions. This work provides a deep learning-based object detection framework for Mars rover navigation that leverages the AI4Mars dataset and deals specifically with terrain elements that are important for navigating safely. A preprocessing pipeline is created to convert semantic segmentation maps into bounding box detections compatible with YOLO across four different terrains: soil, bedrock, sand, and big rock. To make the detection more robust against noisy and incomplete labels, data augmentation techniques and mask filtering techniques were introduced in the process. YOLOv8 architecture was used as the backbone of the framework to extract spatial features and hyperparameters were tuned in a systematic way using iterative optimization methods to improve convergence and overall detection performance. The experimental results suggest that the pipeline is effective at supporting real-time terrain awareness, and demonstrate that reliable perception systems are necessary for improving autonomous navigation on planetary surfaces.