Depth Estimation for Small Obstacles Based on Monocular Vision

In order to solve the problem of small obstacles avoidance in ultra-low altitude flying of unmanned aerial vehicles (UAVs), a depth estimation algorithm based on monocular visual for small tower obstacles is proposed by combining deep learning and traditional methods. With the advantage of convolutional neural networks to extract the salient features of the images, a lightweight and multi-level residual convolution neural network for obstacle segmentation is proposed. The cross-entropy loss function is optimized by adding the weight of the background and obstacles area. A method based on pairs of coplanar points is introduced to complete the depth estimation of a single frame image. Moreover, the recursive formula of sequence images depth results is derived, and a multi-frame voting optimization algorithm is proposed. Finally, the simulation results show that the proposed algorithm can effectively realize the fast segmentation and accurate depth estimation of the small tower obstacles in the sequence images.