Depth map estimation with stereo images
First Claim
Patent Images
1. A method, comprising:
- inputting and processing first and second stereo images captured by stereo cameras at a same time with one or more deep neural network maximum pooling layers;
processing the stereo images with one or more deep neural network upsampling layers;
determining one or more three-dimensional depth maps from the stereo images by solving for stereo disparity in the first and second stereoscopic images with the one or more deep neural network maximum pooling and with the upsampling layers; and
piloting a vehicle based on the one or more depth maps;
wherein the deep neural network maximum pooling layers and the deep neural network upsampling layers are based on cross-correlating a first kernel from the first stereo image with a second kernel from the second stereo image to determine stereo disparity to determine the one or more depth maps;
wherein the deep neural network maximum pooling layers and the deep neural network upsampling layers are trained with training stereo images and associated ground-truth depth maps, wherein the deep neural network maximum pooling layers and the deep neural network upsampling layers are trained with LIDAR data.
1 Assignment
0 Petitions
Accused Products
Abstract
Vehicles can be equipped to operate in both autonomous and occupant piloted mode. While operating in either mode, an array of sensors can be used to pilot the vehicle including stereo cameras and 3D sensors. Stereo camera and 3D sensors can also be employed to assist occupants while piloting vehicles. Deep convolutional neural networks can be employed to determine estimated depth maps from stereo images of scenes in real time for vehicles in autonomous and occupant piloted modes.
38 Citations
14 Claims
-
1. A method, comprising:
-
inputting and processing first and second stereo images captured by stereo cameras at a same time with one or more deep neural network maximum pooling layers; processing the stereo images with one or more deep neural network upsampling layers; determining one or more three-dimensional depth maps from the stereo images by solving for stereo disparity in the first and second stereoscopic images with the one or more deep neural network maximum pooling and with the upsampling layers; and piloting a vehicle based on the one or more depth maps; wherein the deep neural network maximum pooling layers and the deep neural network upsampling layers are based on cross-correlating a first kernel from the first stereo image with a second kernel from the second stereo image to determine stereo disparity to determine the one or more depth maps; wherein the deep neural network maximum pooling layers and the deep neural network upsampling layers are trained with training stereo images and associated ground-truth depth maps, wherein the deep neural network maximum pooling layers and the deep neural network upsampling layers are trained with LIDAR data. - View Dependent Claims (2, 3, 4, 5, 6, 7)
-
-
8. An apparatus, comprising:
-
a processor; and a memory, the memory including instructions to be executed by the processor to; input and process first and second stereo images captured by stereo cameras at a same time with one or more deep neural network maximum pooling layers; process the stereo images with one or more deep neural network upsampling layers; determine one or more depth maps by solving for stereo disparity in the first and second stereoscopic images with the one or more deep neural network maximum pooling and with the upsampling layers; and pilot a vehicle based on the one or more depth maps; wherein the deep neural network maximum pooling layers and the deep neural network upsampling layers are based on cross-correlating a first kernel from the first stereo image with a second kernel from the second stereo image to determine stereo disparity to determine the one or more depth maps; wherein the deep neural network maximum pooling layers and the deep neural network upsampling layers are trained with training stereo images and associated ground-truth depth maps, wherein the deep neural network maximum pooling layers and the deep neural network upsampling layers are trained with LIDAR data. - View Dependent Claims (9, 10, 11, 12, 13, 14)
-
Specification