Development of a system for fully autonomous vehicles in dirt road conditions using artificial intelligence algorithms

Stabilized and dirt roads are transportation routes commonly preferred in rural areas, typically consisting of untreated, natural ground used directly. The automatic detection of such roads and safe navigation upon them are of critical importance, particularly for autonomous ground vehicles. While the majority of recent studies have been based on Unmanned Aerial Vehicles, applications involving ground vehicles face various challenges due to image quality variability caused by environmental factors (lighting conditions, weather, and ground variations). In this study, a unique dataset was created by acquiring 4,867 image frames across different seasons, weather conditions, dust loads, and lighting scenarios using a front-view camera mounted on our custom-developed Utility Task Vehicle platform. The images were annotated pixel-wise as 'road' and 'non-road' using the LabelMe software and used to train a U-Net deep learning architecture. The vehicle is equipped not only for road segmentation but also with the YOLO object detection algorithm to perceive surrounding obstacles. Detecting obstacles on the road such as humans, animals, and rocks, the system ensures safe driving by stopping the vehicle when necessary. Furthermore, a control system was established by integrating stepper motors, linear actuators, GPS, and an accelerometer to enable the vehicle's fully autonomous movement capability. In this way, the vehicle can not only detect the road but also automatically perform steering, acceleration, stopping, and maneuvering operations. Consequently, the developed system has demonstrated its effectiveness and reliability in the automatic detection of stabilized, dirt, and off-road tracks under various conditions; it presents an integrated approach that can be directly implemented for unmanned ground vehicles, agricultural automation, and logistics applications in rural areas. Keywords: Artificial Intelligence, Image Processing, Road Segmentation, U-Net, YOLO, Jetson Nano, Autonomous Ground Vehicle