Research On Depth Estimation Of UAV Monocular Image Based On Deep Learning

Monocular depth estimation(MDE)refers to the process of obtaining a dense depth map from a single RGB image.Since monocular images are easy to acquire by simple sensors,they are widely used in tasks such as self-motion estimation,obstacle avoidance,and scene understanding.The traditional method mainly relies on feature extractor,which can generally achieve high accuracy in a specific environment,but does not perform well in an environment with less texture or low contrast,and requires pre-processing and post-processing,which will increase the computational burden and is not suitable for real-time control of micro-processing platforms such as UAVs.In recent years,deep learning methods have promoted the development of MDE,but most of the deep neural networks designed to improve accuracy are not suitable for micro platforms with high requirements for computing resources and real-time performance.In order to balance accuracy and speed to meet the needs of drones,this paper carries out the following work:(1)Aiming at the problems of complex structure and high network delay of most MDE networks,an efficient codec structure network-MCIN is proposed.After comparing the network structure and performance indicators of several mainstream encoders,an efficient encoder network is selected,and the network uses deep separable convolution as the basic unit,which effectively improves the time required for network coding.At the same time,under the premise of not changing the final output effect,the convolution kernel of the decoder is decomposed,and then the nearest neighbor interpolation with factor 2 is used to increase the resolution of the feature map,so that the final output depth map and the original input image are consistent in size.(2)In view of the over-parameterization problem of monocular depth estimation network proposed in this paper due to the splicing method of network structure and parameter redundancy in the original network,the network pruning method is used to delete the redundant parameters in the depth estimation network proposed in this paper by using the performance performance after network compression as feedback,and then fine-tune the entire network,and finally obtain a lightweight network that can meet the needs by repeating such operations.Experimental results show that compared with traditional methods,the proposed network has a smaller memory and running time within the acceptable range of accuracy loss,and the overall performance is greatly improved.