Research Of Monocular Depth Information Acquisition Method Based On CNN

One of the popular research directions in the field of computer vision is to use image collections or sequences such as two-dimensional images to recover the depth information of objects in the scene.Different from the depth information acquisition method based on binocular vision,monocular stereo vision acquires the scene three-dimensional environment information from the image sequences acquired by the monocular camera,because the image acquisition device used is closer to the daily application,the demand is more extensive,and its research is of great significance to promote the development of computer vision.In the research of depth information acquisition of monocular stereo vision,feature point matching is the key to extracting depth information,and its efficiency and accuracy directly affect the speed and result of depth information acquisition.In order to explore a more suitable method for monocular depth information acquisition and improve the accuracy of the extracting depth information,a monocular depth information acquisition method combining convolutional neural network and local-based stereo matching is designed.Firstly,self-calibration is used to obtain the internal parameters and geometric correction is performed using external geometric constraints,and the pre-processing of specifying two frames of images in the image sequence acquired by the monocular camera is completed.Then the convolutional neural network structure for calculating the matching cost is designed.At the same time,in order to complete the determination of the matching region and the calculation of the matching cost,a matching point data set and network output expression which can be used for network training are constructed.After determining the objective function,the standard data set is used to train the network structure.In order to calculate a more accurate disparity,a cost-based cross method is designed to aggregate the initial matching cost for the discontinuous region in the image,and the semi-global matching algorithm is used to optimize the matching cost after aggregation.Finally,the disparity optimization method is used to calculate the disparity representing the depth information,and the disparity map is obtained.The KITTI data set is processed by the method designed by the project,and the influence of the network structure parameters on the running time and mismatching rate is analyzed.Experiments were carried out according to the standard data and the data taken by a monocular camera.Comparing the error percentage and effect of the running result,it is found that the designed method has a lower mismatch rate,and the contour is clearer and the noise is better removed.The experimental results show that the method of subject design can use the image sequence acquired by the monocular camera to effectively complete the calculation of the matching cost and obtain a clear disparity map.