UAV Obstacle Avoidance by Applying Deep Learning

Abstract

In the last decades, many algorithms in the artificial intelligence field, including machine learning, deep learning, and so on, have been implemented for not only computer vision, but also in mobile vehicles. Thus, to acquire a fast responding and accurate algorithm, this project presents several tests of different deep learning models included in an obstacle avoidance algorithm. In this experiment, a Mavic Air UAV is used for testing this algorithm. The system is designed as two separate parts. One is a pre-operation machine learning system (Keras machine learning framework), the other one is a ground-station control system (a UWP application). CNN, Time-series, semantic segmentation, and instance segmentation models are trained and tested with a custom dataset in this project. The test flight is taken at a parking lot where the main obstacles are the trees and the light pole in the median of the parking lot. The model training results show the multi-stage models outperform the End To End (E2E) models according to various metrics. The test flight shows that even though the model has acceptable performance, the long inference time of the large size of the model limits the flying speed too much to prevent collisions.