Object Tracking Pose Estimation Using Simulated Sensor Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object tracking methods in augmented reality require extensive acquisition and compilation of sensor data for training deep neural networks, which is time-consuming and costly, and may not be feasible in security-relevant applications where detailed object data is confidential.
Innovation Solution
Training a deep neural network using simulated sensor data generated from three-dimensional object models, such as CAD data, to estimate sensor poses without requiring actual sensor data, allowing for high accuracy and confidentiality in object tracking.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If real sensor data is used for training deep neural networks in object tracking, then measurement precision is improved, but loss of time and loss of substance increase due to extensive data acquisition and compilation
Solution Approach 1:
The patent pre-generates comprehensive training datasets using simulated sensor data before the actual training phase. By preparing all necessary training examples in advance through simulation, the system eliminates the need for time-consuming real data acquisition during training, thus resolving the contradiction between measurement precision and loss of time
Solution Approach 2:
The patent creates synthetic copies of real sensor data through simulation. Instead of using actual sensor data for training, the system generates realistic simulated sensor data that replicates the characteristics of real data, thereby maintaining measurement precision while avoiding the time and resource costs of collecting real data
2Measurement precision
If real sensor data is used for training deep neural networks in object tracking, then measurement precision is improved, but loss of substance increases due to extensive data acquisition and compilation
Solution Approach 1:
The patent uses simulated sensor data as synthetic copies替代 real sensor data for training purposes. This approach maintains the statistical properties and characteristics needed for accurate tracking while dramatically reducing the storage resources required, as simulated data can be generated on-demand rather than stored in large datasets
Solution Approach 2:
The patent employs simulation to generate training data with varied parameters and conditions. By changing simulation parameters to create diverse training scenarios, the system achieves comprehensive training coverage without needing to store vast amounts of real sensor data, thus reducing loss of substance while maintaining measurement precision
3Measurement precision
If detailed object data is used for training, then measurement precision is improved, but confidentiality is compromised in security-relevant applications
Solution Approach 1:
The patent uses simulated sensor data that replicates the visual and sensor characteristics of objects without requiring access to the actual objects or their detailed specifications. This synthetic copying approach maintains tracking accuracy while preserving confidentiality, as the simulation process does not need to know or store sensitive object information
Solution Approach 2:
The patent introduces simulation as an intermediary between the training process and the actual objects. The simulator acts as a mediator that generates realistic training data without direct access to confidential object data, thus enabling accurate tracking while maintaining security and confidentiality in sensitive applications
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
The invention relates to a method for providing an object tracking function, wherein, on the basis of detected sensor data (5) of a sensor (20), the object tracking function provides a sensor position (12) relative to an object to be tracked, wherein the object tracking function is provided by means of a deep neural network (1) that has been trained in an object-specific manner, and wherein the deep neural network (1) that has been trained in an object-specific manner is trained using simulated sensor data (8). The invention also relates to a device (30) for data processing, a computer program and a computer-readable storage medium.