Object Depth Measurement and Filtering from Monocular Images for Unmanned Aerial Vehicles

Abstract

The flight safety of low-altitude small fixed-wing unmanned aerial vehicles (UAVs) is often threatened by obstacles such as buildings. This requires UAVs to have the ability to autonomously measure the depth of objects ahead. However, existing depth measurement methods based on multiview geometry and handcrafted features still have problems in accuracy and scene suitability. This paper proposes an object depth measurement and filtering method for UAVs by using monocular images. Firstly, the length of the line segment between feature points instead of the pixel position of feature point is used to solve object depth, which reduces the adverse effect of feature matching error. Meanwhile, in order to adapt to UAV platforms, height and attitude changes are both considered in the modeling process. Moreover, the sequence of object depth values corresponding to the image sequence is filtered by an extended Kalman filter to reduce oscillations. The effectiveness of the whole scheme is verified by visual simulation. Results show that the proposed method achieves better accuracy than other depth measurement methods based on multiview geometry.