Monocular Depth Mapping With Multi-Scale CNN Feature Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for determining depth information from a single image, rather than stereoscopic images, are not reliable for applications like autonomous vehicle control, as they fail to accurately estimate distances of objects from a non-stereoscopic input image.
Innovation Solution
A computer-implemented method using a convolutional neural network with depth map layers that apply multiple scaling filters to an input image, generating scaling maps and selecting the scaling map with the largest pixel value to determine depth information, allowing feature detection across various scales and estimating object distances.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If stereo cameras are used to capture stereo images for depth determination, then measurement precision of depth information is improved, but device complexity increases
Solution Approach 1:
The patent uses a single camera to capture images and creates virtual depth information through computational methods. Instead of using multiple physical cameras (stereo system), the invention copies the depth determination function through software-based scale space analysis and feature matching, replacing complex hardware with computational algorithms that simulate depth perception from monocular images.
Solution Approach 2:
The patent replaces the mechanical/optical stereo camera system with a computational image processing system. The physical mechanism of using two cameras to capture different perspectives is substituted by mathematical operations on single images, including scale space filtering, feature detection, and geometric reasoning to infer depth relationships.
2Measurement precision
If multiple scaling filters are applied to generate scaling maps for depth determination, then measurement precision of object distances is improved, but loss of time in processing increases
Solution Approach 1:
The patent performs scale space filtering and generates scaling maps at different scales before the actual depth determination process. By pre-computing these feature maps at multiple scales, the system prepares the data structure in advance, allowing faster depth estimation during runtime by simply comparing pre-generated scaling maps rather than performing complex computations on the fly.
Solution Approach 2:
The patent divides the image processing into separate scale-specific channels, where each scaling filter processes the image at a specific scale to detect features of particular sizes. This segmentation allows parallel processing of different scale information and enables selective combination of results, improving efficiency by focusing computational resources on relevant scale ranges for the specific application.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a method for determining a depth information image for an input image (B), in order more particularly to control vehicle functions subject to the depth information image, wherein the input image (B) is processed with the aid of a convolutional neural network (2), wherein the convolutional neural network (2) has several layers which process the input image (B) sequentially and in each case convert an input feature map (MKI) into an output feature map (MKO), wherein at least one of the layers is designed as a depth map layer, wherein the depth information image is determined subject to a depth map layer, wherein in the depth map layer (22) - an input feature map (MKI) of the depth map layer (22) is folded with several scaling filters (23, 23a, 23b, 23c) to obtain respective scaling maps (24a, 24b, 24c), - the several scaling maps (24a, 24b, 24c) are compared pixel by pixel in order to generate a corresponding output feature map (MKO) in which every pixel corresponds to a corresponding pixel from a selected scaling map of the scaling maps (24a, 24b, 24c), - a scaling feature map (SK) is generated by assigning each pixel of the scaling feature map (SK) to a piece of information that specifies the scaling map (24a, 24b, 24c) from which the pixel of the output feature map (MKO) is selected; wherein the depth information image corresponds to the scaling feature map (SK) or is determined subject to the scaling feature map (SK).