| Abstract [eng] |
The study investigates monocular depth estimation methods that can be used for real-time autonomous object avoidance systems for unmanned aerial vehicles. After analyzing and experimenting different depth estimation models, the main focus is on the Depth Anything V2 models. In order to evaluate the quality of obstacle avoidance models, several metrics are determined, such as distance estimation errors, collision rate, processing time and resource usage in real time. All of these metrics were evaluated using three types of models based on Depth Anything V2: base, trained on images created in the simulation, and short range models developed in the experiments. Experimental results show that small models focused on short distances achieve more stable depth predictions, lower collision rates and better accuracy in real time compared to larger models that require more time and computing resources for image processing. |
