3D Depth Map Accuracy via Machine Learning Noise Removal
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Time of flight 3D camera mapping systems face accuracy and noise issues due to decreased reflected laser power with increasing distance, leading to increased measurement time and power consumption when trying to improve resolution.
Innovation Solution
Applying machine learning, specifically convolutional neural networks, to process and classify depth map images, reducing the need for high-resolution sensors and increased laser power by enhancing image processing and noise removal.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If laser power is increased to improve depth map resolution, then measurement precision improves, but power consumption increases and safety issues arise
Solution Approach 1:
A machine learning model acts as an intermediary between the low-resolution depth data from the time of flight sensor and the desired high-resolution depth map. The model processes and enhances the input data to produce improved output without requiring increased laser power, thus resolving the contradiction between measurement precision and energy consumption
Solution Approach 2:
The system creates an enhanced copy of the depth map by applying machine learning algorithms that generate high-resolution depth information from low-resolution input. This virtual copying and enhancement process allows the system to achieve high measurement precision without physically increasing laser power or hardware resolution
2Measurement precision
If laser power is increased to improve depth map resolution, then measurement precision improves, but safety issues arise
Solution Approach 1:
The machine learning model serves as a mediator that enables high-resolution depth mapping without requiring high laser power. By processing low-resolution data computationally, the system achieves precise measurements while maintaining safe laser power levels, thus resolving the contradiction between measurement precision and safety
Solution Approach 2:
The system generates a high-resolution copy of the depth map through machine learning enhancement rather than through high-power laser illumination. This computational copying approach allows precise measurements to be obtained without exposing objects or users to harmful laser intensities
3Measurement precision
If measurement time is increased to improve depth map accuracy, then measurement precision improves, but productivity decreases
Solution Approach 1:
The machine learning model is pre-trained on large datasets of depth images and their corresponding high-resolution versions. This preliminary training enables the model to rapidly enhance new depth maps in real-time without requiring extended measurement or processing time, thus resolving the contradiction between accuracy and productivity
Solution Approach 2:
The system creates enhanced depth map copies through pre-trained machine learning models that can process and enhance images rapidly. This approach provides high-accuracy results without requiring multiple repeated measurements or extended processing time, maintaining both precision and productivity
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Improves the accuracy and speed of 3D depth mapping systems by classifying objects and combining 2D and 3D image information, reducing the necessity for high-resolution sensors and increased power consumption.
Implementation Method 1
three dimensional (3D) depth maps may be generated by each device of the other devices or objects within line of sight. Such 3D depth maps are generated typically using 'time of flight' principles, i.e., by timing the periods from laser transmission to reception of each reflection
Implementation Method 2
timing the periods from laser transmission to reception of each reflection
Data Source
AI summary
Machine learning is applied to both 2D images from an infrared imager imaging laser reflections from an object and to the 3D depth map of the object that is generated using the 2D images and time of flight (TOF) information. In this way, the 3D depth map accuracy can be improved without increasing laser power or using high resolution imagers.