Monodepth Prediction Based On Deep Learning

Monocular depth prediction is a research topic that has received much attention in computer vision,and has wide application such as autonomous driving,VR games,and film and television production.In recent years,deep learning has made breakthrough progress in computer vision.However,at present,there are still many problems to be solved in this field.For example,the process of using radar laser to collect depth data is costly and is greatly affected by weather,illumination and other objective factors;the method of depth information based on sparse depth map recovery has edge depth Continuous problems.Stereo images are used for training to improve the prediction accuracy of the depth prediction network.Researches in methods such as deep network structure,stereo matching,and disparity map optimization are essential.The specific work is as follows:1.The DenseNet convolution module is used to repalce the original network modules in the monocular depth prediction framework,which deepens the number of network layers and improve information transformation.Experiments show that the depth prediction framework based on DenseNet modules has less local errors and higher accuracy.2.Cost aggregation based on reconstruction to strengthen the binocular matching constraints is proposed in cost function.Post-processing optimization is also performed on the disparity map outputs.In this thesis,image reconstruction techniques are used at both RGB and disparities to compare the similarity between the reconstructed image and the opposite image of the sampled one.Cost aggregation strengthen the matching constraints of stereo training.At the same time,the smoothing and post-processing method are used to optimize the generated dense disparities,to make disparities smooth without losing the edge of the object.3.The depth prediction model based on KITTI datasets is integrated into a 2D to 3D system,Monocular videos are decomposed into image sequences and uses DIBR to synthesize virtual viewpoint images.Then generated images combines with the original ones to generate 3D images.Evaluates of the model from multiple angles can increase the practicality of the method.Experiments show that the proposed monodepth prediction framework based on DenseNet and cost aggregation based on reconstruction can effectively improve the monocular depth estimation accuracy.Compared with other depth prediction methods,the method proposed in this thesis has a smaller average error rate and the resulting depth map is smoother.