Research And Application Of Scene Depth Estimation Method For Monocular Image With Self-supervised Mechanism

Depth estimation technology can obtain depth information from two-dimensional images,which is of great significance in the field of computer vision.It can help restore the threedimensional structure of the scene,and has a wide range of applications in the fields of augmented reality,autonomous driving and robot vision.In recent years,with the wide application of deep learning and convolutional neural networks,depth estimation algorithms have developed rapidly.However,the data set with high precision depth value is difficult to collect and the equipment is expensive.Therefore,the monocular depth estimation algorithm using self-supervision mechanism comes into being,which aims to give up the difficult depth data and use image reconstruction to complete the depth estimation task.But at the same time,the realistic factors such as no texture,occlusion and illumination transformation also bring great challenges to the self-supervised monocular depth estimation.This paper mainly studies the self-supervised monocular depth estimation based on video sequence,designs and deploies the vehicle anti-collision system based on monocular depth estimation.The main research achievements are as follows:(1)Self-supervised monocular depth estimation based on full-scale feature fusion.In order to solve the problem that U-codec networks only use the jump connection mode for feature fusion at the same scale,resulting in inadequate feature fusion at each scale,a self-supervised monocular depth estimation network based on full-scale feature fusion is proposed.In the process of decoding,the high resolution and the same resolution features obtained by the encoder are fused with the low resolution features obtained by the previous decoder and the inverse depth map of the upper level in the way of channel stacking,so that the obtained features contain both global information and local information.The weights of different scale feature map fusion were redesigned to make the fusion strategy more reasonable and effective.In this design,a chain residual module is used to obtain context information from the fused features,supplement and optimize the feature information,and finally realize the full-scale feature fusion,thus significantly improving the accuracy of depth estimation.The model was validated on KITTI and Make3 D datasets,where the absolute error of depth values was reduced to 1.06 and the accuracy of threshold values of 1.25 was improved to 0.89.(2)Self-supervised monocular depth estimation based on parallel multiresolution.A selfsupervised monocular depth estimation network based on parallel resolution is proposed to solve the problem of information loss and image quality degradation in U-codec networks during constant up-down sampling.Firstly,a parallel multi-resolution feature extraction module is designed to extract the features of the input image.By maintaining the feature map with the highest resolution,the upper and lower sampling of the image is prevented from causing information loss.At the same time,the feature information with different resolutions is exchanged at each stage.The feature adaptive fusion module is designed to integrate the feature maps from the highest resolution at different stages and complement each other.Finally,the channel attention module adjusts the weights of different channels in the feature map and uses the optimized feature information for depth estimation to improve the accuracy of the depth estimation results.The model was validated on KITTI and Make3 D datassets,where the absolute error of depth values was reduced to 0.099 and the accuracy was improved to 0.902 with a threshold of 1.25.(3)Vehicle anti-collision system based on depth estimation.The self-supervised monocular depth estimation method is applied in practice,and the above parallel multiresolution monocular depth estimation network is packaged,and the vehicle anti-collision system based on depth estimation is constructed.Its main functions include obstacle detection,real-time depth estimation and anti-collision alarm.Obstacle detection is to use YOLOv3 algorithm to detect common obstacles on the road in the video.Rectangular boxes are used to divide and mark the obstacle categories.Real-time depth estimation is to use the depth estimation network to estimate the depth of the video scene at the pixel level in real time to obtain the depth information of the distance between vehicles and obstacles.Anti-collision alarm will determine whether the obstacle is too close to the vehicle,and trigger the alarm if the distance is lower than the threshold.Then model pruning technology is used to cut the depth estimation model to meet the real-time requirements of the system operation.And through Py Qt5(Qt Designer)design system graphical interface,visual real-time video and running results.The demonstration results show that the real-time performance of the system is good and has certain application value.Finally,the deficiencies of the system are analyzed and the corresponding improvement methods are proposed.