Obstacle avoidance for a UAV via a neural network method
Zhaokai Ma · DR-NTU (Nanyang Technological University) · 2026
Unmanned Aerial Vehicles (UAVs) are not an emerging domain, but have received more attention from researchers in recent years mainly by their industrial and military applications. More techniques are adopted in tasks involving autonomous navigation, while still facing challenges in avoiding collisions, particularly in reacting to objects in unknown and dynamic environments. This research focuses on obstacle avoidance using visual input alone and being reactive to the potential threat. To achieve this task, this dissertation employs a neural network pipeline, combining collision detection with directional escape estimation. The system leverages CNN and LSTM for decision making, and integrates optical flow estimation with clustering to guide escape direction prediction. Experiments are conducted on two UAV datasets to validate effective performance in both detection accuracy and directional prediction quality. The comprehensive experiments and detailed analysis may provide some guidance for future explorations.