The Recovery Of Target Distance From A Single Gray-scale Image

In the field of computer vision,depth estimation and 3D reconstruction are important research directions.The key is to recover the distance from the target in the image to the camera.This paper mainly studies the recovery of target distance from a single gray-scale image,and studies it from two aspects of visible gray images and infrared gray images.Firstly,for visible light images,monocular depth estimation is performed using supervised learning.A novel upsampling layer is introduced,and this paper implemented the interleaving operation of it.This paper adopts the architecture constructionof fully convolutional neural network,including encoder network and decoder network: the residual neural network is used as the encoder network to extract feature maps from input the image,and then the upsampling layer is used to construct the decoder network to gradually increase the size of the feature maps,thereby outputting the predicted depth map.On this basis,this paper proposes two models,a u-shaped fully convolutional residual network and a fully convolutional dilated residual network,which are meant to recover target distance in gray-scale images.The u-shaped fully convolutional residual network proposed in this paper uses skip connections to concatenate feature maps in the model near the input with those near the output.Those skip connections can not only fuse context features and retain more edge information,but also enhance information transmission and speed up network convergence.The fully convolutional dilated residual network introduces an dilated residual network as its encoding part to extract image features.Without increasing the number of weight parameter or reducing the feature size,the encoding part can still maintain the receptive field size of neurons in the subsequent network layers,so that the features containing more spatial information is extracted and transmitted to the subsequent upsampling layers,which makes the model better learn the mapping relationship from a pixel to the real depth value and enhances the fitting ability of the model.Secondly,the models are evaluated and analyzed on the NYU Depth Dataset V2,which shows that the proposed methods can effectively recover depth information from a single visible image and are superior to some mainstream methods in terms of accuracy and error.Subsequently,the models are applied to recover the target distance from single visible light gray image,and the validity of the models to recover the target disdance is verified by experiments.Finally,a thermal imager is used to acquire the infrared image sequence of the same target at different distances,and the true distance of the target is recorded.And the a target detection algorithm of infrared gray image is realized by means of binarization,morphological opening and closing operations and calculation of connected domains,and a monocular vision ranging method for infrared image targets is realized by using the geometric relationship between the image coordinate system and the road coordinate system.The result of the experiment shows that when the front target distance is 7.5 meters to 15 meters,the target can be detected well,and the relative errors of the distance prediction results are within 11%.