The subject of the research is the problem of preventing collisions of unmanned aerial vehicles based on the analysis of visual information obtained from onboard cameras. The focus is on the development and evaluation of algorithms that allow the drone to timely recognize obstacles and calculate the time to collision (Time-to-Collision, TTC). The author thoroughly examines traditional methods, including monocular algorithms based on optical flow, stereo vision, temporal accumulation approaches, as well as neural network solutions. Special attention is given to their strengths and weaknesses: limited detection range, sensitivity to texture and lighting, computational costs, and resilience to false alarms. The object of the study is a new combined method DUO-TTC (Dual-Camera, Uncertainty-Aware Optical-Flow Time-to-Collision), which combines the advantages of stereo vision and optical flow with multi-frame fusion, taking into account the uncertainty level of the data. The work aims to enhance the range of 'vision' for drones and the reliability of early obstacle detection while maintaining low computational load. The research methodology is based on the development and modeling of the DUO-TTC algorithm, which combines the use of two cameras, optical flow analysis, and multi-frame triangulation. The algorithm assesses the uncertainty of each channel, providing more reliable and long-range obstacle detection compared to baseline methods. The main findings of the study indicate that the proposed DUO-TTC method significantly outperforms classical approaches (Mono-TTC, Stereo, TemporalStereo) across key performance metrics. It nearly doubles the rate of timely warnings, extends the obstacle detection range to 30-40 meters, and maintains a low level of false alarms. A notable contribution by the author is the demonstration that combining various perception channels and accounting for uncertainty can compensate for the shortcomings of individual methods and ensure more stable performance in challenging conditions – in low light, image blur, and lack of textures. The novelty of the research lies in the integration of stereo vision, optical flow, and temporal analysis into a unified system optimized for the limited computational resources of drones. This opens up prospects for creating more reliable autonomous navigation systems capable of safely functioning in real flight conditions.
Konnikov et al. (Tue,) studied this question.