Motion Estimation Circuit Using Deep Neural Networks for SLAM
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current motion estimation techniques in Visual SLAM systems face challenges in accurately estimating camera motion and creating reliable environmental maps, particularly in environments with complex features and dynamic conditions.
Innovation Solution
A motion estimation device and method that includes a dedicated circuit for processing input frames from cameras, utilizing deep neural networks to estimate pixel distances and reliability, detect feature points, and perform matching and mapping operations, thereby enhancing the accuracy of motion estimation and map creation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural networks are used to estimate pixel distances and reliability for each pixel, then measurement precision of distance is improved, but device complexity increases
Solution Approach 1:
The processing circuit is segmented into multiple functional units: a deep neural network for distance estimation, a reliability estimation unit, a feature point detection unit, and a motion estimation unit. Each unit handles a specific task in the motion estimation pipeline, allowing complex processing to be distributed across specialized components rather than concentrated in a single complex block.
Solution Approach 2:
The system performs preliminary distance estimation and reliability estimation for all pixels before feature point detection. By pre-calculating distance and reliability information, the system prepares data in advance that guides subsequent feature point selection and matching, reducing the computational burden during critical motion estimation phases.
2Measurement precision
If feature points are selected based on reliability estimation, then motion estimation accuracy is improved, but processing time increases
Solution Approach 1:
The system applies reliability estimation selectively to identify high-quality feature points rather than uniformly processing all pixels. By focusing computational resources on regions with high reliability scores, the system achieves accurate motion estimation while avoiding unnecessary processing of low-quality regions.
Solution Approach 2:
The system performs reliability estimation for all pixels but only proceeds with detailed feature point matching for those exceeding a reliability threshold. This partial action approach ensures that time-consuming matching operations are applied only to promising candidates, balancing accuracy with processing efficiency.
3Reliability
If moving points are excluded from feature point matching, then reliability of map creation is improved, but device complexity increases
Solution Approach 1:
The system extracts and identifies moving points from the scene using the deep neural network's distance estimation capabilities. By separating moving objects from the static environment, the system can exclude them from map creation processes, ensuring that the environmental map represents only stable, reliable features.
Solution Approach 2:
The reliability estimation unit acts as an intermediary between distance estimation and feature point matching. It filters out unreliable features including moving points before they enter the motion estimation pipeline, preventing them from degrading map creation reliability without requiring complex real-time tracking of individual moving objects.
Data Source
AI summary
According to one embodiment, a motion estimation device includes a first receiving circuit that receives a first input frame and a calculation circuit that performs motion estimation processing on the first input frame. The calculation circuit estimates a distance for each pixel of the first input frame and estimates a reliability of the distance for each pixel of the first input frame based on pixel information of the first input frame.