Optical Flow Determination Using CNN Spatial Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining optical flow in images are either computationally expensive and accurate or inexpensive but less accurate, failing to effectively address the need for efficient and precise motion analysis in computer technology.
Innovation Solution
The use of a convolutional neural network (CNN) to process video frames, generating feature maps and employing a spatial matching layer to determine optical flow by averaging displacement measurements and utilizing histograms to predict pixel movement along axes, with the network trained using optical flow training data to minimize inaccuracies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional optical flow algorithms are used, then measurement precision is improved, but computing resource consumption increases
Solution Approach 1:
The patent replaces traditional mechanical/optical computational methods with a neural network-based system. The neural network learns optical flow patterns through training data, substituting explicit pixel-by-pixel computational algorithms with a trained model that can predict optical flow more efficiently while maintaining accuracy.
Solution Approach 2:
The patent performs preliminary action by training the neural network offline using extensive optical flow training data before deployment. This pre-training phase captures complex motion patterns in advance, allowing the network to make accurate predictions during runtime without requiring heavy computational resources for real-time processing.
2Productivity
If fast optical flow algorithms are used, then productivity is improved, but measurement precision deteriorates
Solution Approach 1:
The patent changes the fundamental parameters of the optical flow computation by using a neural network with learnable weights and biases instead of fixed algorithmic parameters. The network architecture, activation functions, and training objectives are all parameters that can be optimized to achieve both speed and accuracy simultaneously.
Solution Approach 2:
The patent substitutes traditional iterative optimization algorithms with a neural network inference process. Once trained, the network performs parallel computations across the entire image, achieving high processing speeds while maintaining the precision benefits of more complex algorithms through its learned representations.
Data Source
AI summary
Systems, methods, and non-transitory computer-readable media can obtain a first video frame and a second video frame. The first video frame can be processed using a convolutional neural network to output a first set of feature maps. The second video frame can be processed using the convolutional neural network to output a second set of feature maps. The first set of feature maps and the second set of feature maps can be processed using a spatial matching layer of the convolutional neural network to determine an optical flow for at least one pixel.