Monocular Camera Depth Estimation Using Convolutional Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing electronic systems face challenges in estimating depth from monocular images efficiently, as existing methods like stereo cameras and LIDAR systems are costly and impractical for portable devices.
Innovation Solution
A system using a convolution neural network (CNN) trained on monocular RGB images to generate depth perception estimates, with a secondary CNN for confidence metrics, allowing for accurate distance estimation between a mobile device and obstacles without the need for multiple cameras or expensive sensors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If stereo cameras or LIDAR systems are used to determine depth, then measurement precision is improved, but device complexity and cost increase
Solution Approach 1:
The patent replaces complex mechanical/optical depth sensing systems (stereo cameras, LIDAR) with a computational approach using a single monocular camera combined with a convolutional neural network. The CNN processes monocular images to estimate depth by learning features from training data, substituting hardware complexity with software-based depth estimation while maintaining reasonable accuracy for mobile device applications
Solution Approach 2:
The patent uses a single camera to capture images that are then processed by a neural network to create a depth map representation. Instead of using multiple physical sensors, the system creates a computational copy of depth information through image processing and machine learning, allowing depth estimation from standard monocular imagery
2Measurement precision
If stereo cameras or LIDAR systems are used to determine depth, then measurement precision is improved, but device cost increases
Solution Approach 1:
The patent employs a standard monocular camera, which is significantly cheaper than stereo camera systems or LIDAR sensors, combined with a trained neural network model. This approach uses inexpensive, readily available components that can be manufactured at scale for mobile devices, eliminating the need for costly specialized depth sensing hardware while providing sufficient depth estimation capability
Solution Approach 2:
The patent replaces expensive mechanical/optical depth sensing hardware with a computational solution using a single camera and neural network. This substitution dramatically reduces bill of materials costs while maintaining acceptable depth estimation performance for navigation and obstacle detection applications in mobile devices
3Device complexity
If monocular images are used for depth estimation, then device complexity is reduced, but measurement precision deteriorates
Solution Approach 1:
The patent transforms the monocular image data through a convolutional neural network that learns to extract depth-relevant features and parameters. The CNN processes the two-dimensional image to generate a one-dimensional array of depth perception estimates with angle-distance pairs, effectively changing the parameter representation from simple pixel intensities to structured depth information that compensates for the limitations of monocular sensing
Solution Approach 2:
The patent performs preliminary training of the convolutional neural network using labeled depth data before deployment. The network is pre-trained to recognize depth cues and patterns in monocular images, allowing it to provide reasonably accurate depth estimates when deployed in the mobile device. This preliminary training phase enables the system to overcome the inherent limitations of monocular depth estimation through learned feature extraction
Data Source
AI summary
A system of determining a distance between a mobile electronic device and an obstacle in an environment includes an electronic device, and a computer-readable storage medium. The computer-readable storage medium includes one or more programming instructions that, when executed, cause the electronic device to obtain from a camera of a mobile electronic device one or more images of one or more obstacles in a surrounding environment, and apply a convolution neural network to one or more of the obtained images to generate a one-dimensional array of depth perception estimates. The array includes one or more angle-distance pairs. an angle value of each angle-distance pair represents an angle of one or more of the obstacles relative to the camera. A distance value of each angle-distance pairs represents an estimated distance between the camera and one or more of the obstacles at the corresponding angle of the angle-distance pair.