Research And Application Of Monocular Depth Estimation For Dynamic Scenes

In recent years,with the development of deep learning technology and hardware performance,intelligent equipment based on deep learning has become increasingly mature,including depth estimation for ranging,tracking and identification.As a common calculation method in monocular SLAM,unsupervised monocular depth algorithm uses a camera as the main data input source,with low hardware requirements,small computation and fast computing speed.At present,due to the problems of models and the limitations of scenarios,it is often impossible to estimate the depth of specific scenarios effectively.One of the big challenges is scenes that contain dynamic objects.The monocular depth estimation model is based on static scenario assumption,which can not be satisfied with complex application scenarios.However,in theory,the detection rate and accuracy of current highperformance GPU devices have reached a very high level,and the complex network structure of various algorithms and massive training data also provide strong support.Therefore,in view of the inherent problems of the model and the characteristics of the application scene,this paper proposes a brand new network model combined with image segmentation.The specific research work is as follows:Firstly,the projection relation between different frames is deduced for monocular depth estimation,and the modeling principle of depth estimation model is obtained.Then static scene in Kitti data sets collected data set and data set contains dynamic target scene,with the classic SFMLearner network training respectively under the two data sets and analysis of test results,it is concluded that the problems existing in the model under the dynamic target scene,combined with the framework model,after analyzing the characteristics of moving target,put forward the improvement scheme.Aiming at the dynamic object scene,the network structure is designed.Besides the original depth estimation network and selfmotion estimation network,the moving object estimation network is added.Combined with image segmentation,each moving object in the image is covered by mask and then sent to the network for motion estimation.The network structure is the same as the camera selfmotion estimation network.Then train and test the data set containing dynamic target scenes in the previous chapter,compare with SFMLearner network,analyze each evaluation index,and finally verify that the model network in this paper has a better depth estimation effect on dynamic scenes than SFMLearner network.Aiming at the dynamic object scene design model loss function,aiming at the improvement scheme of the second chapter,this chapter improves the SFMLearner network luminosity loss function,introduces a new continuous scale loss function and smooth loss function.For the data set containing dynamic target scene,the weights of two parameters in the luminosity loss function are firstly determined by experiments.Then the weights of all three loss functions are iterated to achieve optimal results.Then the influence of each loss function on the model accuracy was discussed by using ablation experiment.The model proposed in this paper is used for transfer learning.Ship data sets are made and the segmentation network is trained according to the characteristics of sea surface scenarios and unmanned ship targets Then the hardware and experimental scheme of monocular depth estimation system are designed.In the experiment,the depth effect map of unmanned ship is obtained and the calculation distance is obtained by using the idea of normalization.Finally,the measurement results are compared with other ranging schemes,which can verify that the method in this paper can still achieve good results when moving training ship targets.