Object Pose Generation Using Perspective Transformation Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vehicle systems struggle to accurately compute the pose of static or moving objects using camera sensors, requiring significant computational resources and relying on geometry-based computations.
Innovation Solution
The use of an auxiliary network, such as a multilayer perceptron, to generate parameters that enable the transformation of scenes from a first perspective to a second perspective, such as an overhead view, with reduced computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If geometry-based computations are used to compute object pose from camera sensor data, then measurement precision is improved, but device complexity and computational resources increase
Solution Approach 1:
The patent replaces traditional geometry-based computational methods with a trained neural network model. The network is trained offline to learn the mapping from camera sensor data to object pose parameters, substituting complex real-time geometric computations with a trained model that performs the same function with reduced computational burden during deployment.
Solution Approach 2:
The neural network is trained in advance using labeled training data that includes images and corresponding object pose information. This preliminary training phase computes and stores the optimal parameters and weights that encode geometric relationships, so that during actual operation, the network can directly apply these pre-computed parameters without performing complex geometric calculations in real-time.
2Measurement precision
If traditional pose computation methods are used, then measurement precision is maintained, but productivity decreases due to high processing requirements
Solution Approach 1:
Complex geometric computation algorithms are replaced with a trained neural network model that has learned pose estimation patterns during offline training. The network processes camera input data through learned feature extractions and transformations, achieving comparable or superior accuracy with significantly reduced computational complexity and faster processing speeds.
Solution Approach 2:
The patent transforms the pose computation problem from explicit geometric parameter calculations to implicit parameter learning through neural network training. The network learns optimal parameter transformations during training, including coordinate system transformations, perspective projections, and pose parameter extractions, enabling efficient real-time computation without explicit geometric formula evaluations.
Data Source
AI summary
A computer includes a processor and a memory, and the memory stores instructions executable by the processor to extract a feature of an object viewable in a scene captured by a camera at a first perspective. The instructions can also be to transform the scene from the first perspective to a second perspective based on a parameter generated responsive to (a) a reconstruction of the scene from the first perspective based on the scene from the second perspective, and (b) an adjustment of the parameter to minimize differences between the extracted feature and a representation of the extracted feature reconstructed based on the second perspective. The instructions can also be to determine a depth of the extracted feature with respect to the camera based on the parameter and to determine the pose of the object based on the determined depth.