Optical Flow Rectification for Neural Network Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks used in vehicle driving assistance systems are less effective in detecting objects when the vehicle's motion includes components other than longitudinal translation, such as lateral translation, vertical translation, or rotations, due to the limited training data.
Innovation Solution
An improved process for detecting objects using a neural network that includes estimating ego-motion parameters and applying rotation matrices to rectify optical flow maps, converting them into rectified optical flow maps that ignore parasitic motions, and using these rectified maps for training and detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If neural networks are trained with optical flow maps corresponding to longitudinal translation only, then the neural network achieves high detection accuracy for objects during straight-line vehicle motion, but the detection accuracy deteriorates when the vehicle experiences lateral translation, vertical translation, or rotations
Solution Approach 1:
The patent applies parameter changes by transforming the optical flow maps through rotation matrices that adjust the coordinate system based on estimated ego-motion parameters. This converts optical flow maps corresponding to complex motions (lateral translation, vertical translation, rotations) into equivalent optical flow maps representing only longitudinal translation, thereby maintaining detection accuracy across diverse vehicle motion conditions.
Solution Approach 2:
The patent introduces an intermediary transformation process involving rotation matrices and coordinate system adjustments. This intermediary step rectifies the optical flow maps by compensating for parasitic motions through estimated ego-motion parameters, serving as a bridge between the raw optical flow data and the neural network's expected input format.
2Productivity
If the neural network is supplied with raw optical flow maps containing parasitic motions, then the system can process real-time data directly from the image acquisition device, but the detection performance deteriorates due to biases introduced by non-linear vehicle motions
Solution Approach 1:
The patent applies preliminary action by performing the rectification of optical flow maps before feeding them to the neural network. The ego-motion parameters are estimated and rotation matrices are computed in advance to correct the optical flow data, removing biases from parasitic motions before the detection process begins.
Solution Approach 2:
The patent replaces direct mechanical processing of raw optical flow data with a computational geometric transformation approach. Instead of processing the raw optical flow maps directly, the system substitutes them with rectified optical flow maps that have been mathematically transformed to eliminate parasitic motion effects.
3Measurement precision
If the system estimates and rectifies ego-motion parameters to remove parasitic motions, then the neural network achieves improved detection accuracy under various vehicle motions, but the computational complexity and processing time increase
Solution Approach 1:
The patent changes the parameters of the optical flow maps through rotation transformations based on estimated ego-motion parameters. By adjusting the coordinate system and applying rotation matrices, the system transforms complex motion patterns into simplified longitudinal translation representations, reducing the computational burden on the neural network.
Solution Approach 2:
The patent extracts and removes the parasitic motion components from the optical flow maps through the rectification process. By separating and eliminating the effects of lateral translation, vertical translation, and rotations, the system isolates the relevant longitudinal motion information for accurate object detection.
Data Source
AI summary
A process for detecting objects by a neural network. The neural network being supplied at input with at least one original first optical flow map which represents the computerized tracking of moving objects in a scene by analyzing the differences in content between a first image, captured by an image acquisition device in a first position at an earlier time, and a successive second image, captured by the image acquisition device in a second position at a current time. The process includes rectifying the original optical flow map by using the ego-motion estimation information of the image acquisition device.